BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:iCalendar-Ruby
X-WR-CALNAME:UCLA Statistics Seminars
BEGIN:VEVENT
X-SEMINAR-ID:334
SEQUENCE:0
DTEND:20080501T170000
UID:2009-11-17T20:55:58-08:00_282977663@limen.stat.ucla.edu
DESCRIPTION:Education reforms in the last fifteen years have enlivened the 
 teaching of introductory statistics with fewer lectures and more active lea
 rning\, fewer recipes and more conceptual thinking\, fewer contrivances and
  more real data.  However\, bringing these reforms to large multi-section i
 ntroductory courses has been a difficult challenge. Thus\, the buffet model
  was developed at The Ohio State University to use class size as a strength
  rather than a weakness\, to optimize learning for the individual rather th
 an norming for the group\, and to integrate technology as an efficient tool
  rather than an expensive add-on. \n\n Students learn in different ways so\
 , in the buffet model\, different course sections are geared toward differe
 nt learning styles and students are offered a choice of interchangeable pat
 hs to learn the same course objectives.  In order to promote student commit
 ment to follow through on their choices and to enable efficient tracking of
  each student's progress through the course\, the choice of learning modes 
 is exercised through an on-line "contract" entered into by students at the 
 beginning of the quarter.  Students can make an informed choice based on th
 e results of their own learning styles inventory and by reading testimony f
 rom previous students most like themselves.  The buffet structure has been 
 successful in increasing both student satisfaction and student learning. Fo
 r example\, scores on common exams have increased by about a half-letter gr
 ade while dropouts and students needing to retake this required course have
  decreased by about 40%. Finally\, key elements of the buffet strategy can 
 also be adapted to smaller classes to improve student learning.
SUMMARY:Cooking for the Buffet: Individualizing Course Content to Improve L
 earning
DTSTART:20080501T160000
DTSTAMP:20091117T205558
LOCATION:5137 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:4
SEQUENCE:0
DTEND:20090512T160000
UID:2009-11-17T20:55:58-08:00_833377135@limen.stat.ucla.edu
DESCRIPTION:We explore statistical frameworks for the simultaneous\, unsupe
 rvised segmentation and discovery of visual object categories from image da
 tabases. Examining a large set of manually segmented scenes\, we show that 
 object frequencies and segment sizes both follow power law distributions\, 
 which are poorly captured by standard approaches. Motivated by this\, we de
 velop an alternative family of models based on the Pitman-Yor (PY) process\
 , a generalization of the Dirichlet process. This nonparametric prior distr
 ibution leads to learning algorithms which discover an unknown set of objec
 ts\, and segmentation methods which automatically adapt their resolution to
  each image. Generalizing previous applications of PY priors\, we use non-M
 arkov Gaussian processes to infer spatially contiguous segments which respe
 ct image boundaries. Using a novel family of variational approximations\, o
 ur approach produces segmentations which compare favorably to state-of-the-
 art methods\, while simultaneously discovering categories shared among natu
 ral scenes.
SUMMARY:Shared Segmentation of Natural Scenes using Dependent Pitman-Yor Pr
 ocesses
DTSTART:20090512T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:2
SEQUENCE:0
DTEND:20090526T160000
UID:2009-11-17T20:55:58-08:00_466871031@limen.stat.ucla.edu
DESCRIPTION:The talk will survey various statistical techniques for evaluat
 ing the goodness-of-fit of point process models\, such as those used in ear
 thquake forecasting. Commonly used methods involve examining spatial aggreg
 ates over pre-specified grid cells as described in detail by Baddeley\, Tur
 ner\, Møller and Hazelton (2005). These methods can have severe drawbacks\,
  such as loss of power when large grid cells are used\, and enormous varian
 ce in the case of small grid cells. In the case of assessing spatial-tempor
 al models for phenomena such as earthquakes\, alternative diagnostics may b
 e preferable\, such as rescaled and thinned residuals and weighted second-o
 rder statistics. These residual techniques will be described and illustrate
 d here\, and their pros and cons for the purpose of assessing models for ea
 rthquake forecasting will be discussed.
SUMMARY:Goodness-of-fit Testing for Point Process Models for Earthquake Occ
 urrences
DTSTART:20090526T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:9
SEQUENCE:0
DTEND:20090407T160000
UID:2009-11-17T20:55:58-08:00_252329917@limen.stat.ucla.edu
DESCRIPTION:Most previous statistical theories of human causal learning hav
 e focused on learning from summarized contingency data based on binary vari
 ables. The computational theory described here instead provides a trial-by-
 trial model of learning from sequential data. For nonverbal animals\, there
  is no obvious way to present summarized data\; often\, humans also must le
 arn from sequential data in a dynamic environment. In particular\, sequenti
 al models are required to account for influences of the order of data prese
 ntation. A computational theory should enable beliefs to be dynamically upd
 ated by integrating prior beliefs with new observations in a trial-by-trial
  manner. I will present two studies of sequential learning in the context o
 f causal reasoning and memory to illustrate the use of dynamic learning mod
 els in human cognition.
SUMMARY:Learning from Sequential Data
DTSTART:20090407T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:11
SEQUENCE:0
DTEND:20090317T160000
UID:2009-11-17T20:55:58-08:00_601163498@limen.stat.ucla.edu
DESCRIPTION:In this talk I will discuss our winning entries into the 2007 a
 nd 2008 PASCAL Visual Object Category challenges. I will discuss some thoug
 hts on why sliding window detectors often yield competitive object detector
 s\, even when implemented with a simple linear classifier. Our work propose
 s an extension of linear classification that incorporates a search over lat
 ent variables during the classification process. This latent variable model
  allows for a natural way to build in invariance during the recognition pro
 cess\, such as invariance to pose and geometric deformations. We finally de
 scribe large-scale learning methods for training this models from semi-supe
 rvised data. This talk discusses joint work with Pedro Felzenszwalb and Dav
 id McAllester and Ross Girshick.
SUMMARY:Discriminative Models for Finding Objects in Images
DTSTART:20090317T150000
DTSTAMP:20091117T205558
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:418
SEQUENCE:0
DTEND:20091103T160000
UID:2009-11-17T20:55:58-08:00_289454048@limen.stat.ucla.edu
DESCRIPTION:Symmetry is an essential mathematical concept\, as well as a ub
 iquitous\, observable phenomenon in nature\, science and art. Either by evo
 lution or by design\, symmetry implies an efficiency coding that makes it u
 niversally appealing\, especially so to computational science. Recognition 
 and categorization of symmetry and regularity is the first step towards cap
 turing the essential skeleton of a real world problem\, while at the same t
 ime minimizing computational redundancy. However\, symmetry group detection
  from real world data turns out to be a challenging problem that has been p
 uzzling computer vision\, computer graphics and psychology researchers for 
 decades. We explore a formal and computational characterization of real wor
 ld regularity using a hierarchical model of symmetry groups as a theoretica
 l basis\, embedded in a well-defined Bayesian framework. Such a formalizati
 on simultaneously facilitates (1) a robust and comprehensive algorithmic tr
 eatment of the whole regularity spectrum\, from regular (perfect symmetry)\
 , near-regular (approximate symmetry)\, to various types of irregularities\
 ; (2) an effective detection scheme for real world symmetries and symmetry 
 groups\; and (3) a set of computational bases for measuring and discriminat
 ing quantified regularities on diverse data sets.  Besides some theoretical
  background on crystallographic groups in particular\, I shall illustrate v
 arious applications of computational symmetry in texture synthesis\, analys
 is\, tracking\, and manipulation\; human gait and activity recognition\; sy
 mmetry-based dance analysis\; grid-cell clustering\; automatic geo-tagging\
 ; and image ‘de-fencing’. \n\n \n\n \n\n \n\n \n\n \n\n \n\n \n\n \n\n Bios
 ketch \n\n Yanxi Liu received her B.S. degree in physics/electrical enginee
 ring in Beijing and her Ph.D. degree in computer science for group theory a
 pplications in robotics from University of Massachusetts (Amherst). Her pos
 tdoctoral training was at LIFIA/IMAG (France). She also spent one year at D
 IMACS (NSF center for Discrete Mathematics and Theoretical Computer Science
 ) under an NSF research-education fellowship award. Dr. Liu was with the re
 search faculty in the Robotics Institute (RI) of Carnegie Mellon University
  before she joined the Computer Science Engineering and Electrical Engineer
 ing departments of Penn State University in Fall of 2006 as a tenured facul
 ty and the co-director of the lab for perception\, action and cognition (LP
 AC). Dr. Liu's research interests span a wide range of applications includi
 ng computer vision\, computer graphics\, robotics\, human perception and co
 mputer aided diagnosis in medicine\, with two main themes: computational sy
 mmetry/regularity and discriminative subspace learning. Dr. Liu chaired the
  First International Workshop on Computer Vision for Biomedical Image Appli
 cations (CVBIA) in conjunction with ICCV 2005. Dr. Liu served as an area ch
 air or organizing committee member for CVPR08/MICCAI08/CVPR09\, and has ser
 ved as a multi-year chartered study section member for the US National Inst
 itute of Health (NIH). Dr. Liu is a senior member of IEEE and the IEEE Comp
 uter Society.
SUMMARY:Computational Regularity
DTSTART:20091103T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:397
SEQUENCE:0
DTEND:19960226T160000
UID:2009-11-17T20:55:58-08:00_760534849@limen.stat.ucla.edu
DESCRIPTION:Texture is a powerful cue in visual perception\, and texture an
 alysis and synthesis has been an active research area in computer vision. W
 e present a statistical theory for texture modeling and random field approx
 imation\, which combines multi-channel filtering and random field modeling 
 via the maximum entropy principle. Our theory characterizes a texture by a 
 random field\, the modeling of which consists of two steps. (I) Feature ext
 raction: a set of finite-support filters (usually linear) is applied to the
  random field and the marginal distributions of the filter responses are co
 nsidered as features concerning textural appearances\, which is supported b
 y physiological observations on visual cells. (II) Feature fusion: the mode
 l is then derived by maximizing the entropy over all distributions with the
  same marginal distributions as in (I)\, and the resulting model is called 
 Filter\, Random field\, And Maximum Entropy (FRAME) model. Sieve-MLE is use
 d for non-parametric fitting\, where a histogram matching algorithm is prop
 osed for computation and simulation. We think of FRAME as low-dimensional a
 pproximation to the "true" distribution that generates texture images\, and
  prove that any MRF can be approximated arbitrarily close by FRAME. Motivat
 ed by a mini-max entropy principle\, we propose a filter pursuit procedure 
 for selecting a parsimonious set of meaningful filter "words" from a well d
 esigned filter "vocabulary" when modeling a certain texture image. A variet
 y of texture synthesis experiments are described to illustrate our theory (
 intersting pictures will be displayed). Many previous methods and concepts 
 are interpreted and clarified in a unified point of view. \n\n This talk is
  based on the joint work with S.C. Zhu and D.B. Mumford. Support from Prof.
  D.B. Rubin and many helpful discussions with Prof. A.P. Dempster are warml
 y acknowledged.
SUMMARY:A Theory for Texture Modeling and Random Field Approximation
DTSTART:19960226T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:6
SEQUENCE:0
DTEND:20090428T160000
UID:2009-11-17T20:55:58-08:00_542277279@limen.stat.ucla.edu
DESCRIPTION:This talk presents an interactive method for painterly renderin
 g of abstract arts from photos. Psychological studies [Berlyne 1971] sugges
 ted that abstract arts are often characterized by their greater perceptual 
 ambiguities than photos\, and the increased ambiguities invoke more mental 
 efforts for interpretation with aesthetic pleasures. Given an image\, we in
 crease the ambiguities of selected image components (e.g. objects\, regions
 ) interactively to various levels using stochastic image operators\, so tha
 t some components become unrecognizable individually\, but with efforts mea
 surable by delayed response time\, they are still identifiable through info
 rmation propagated from their surrounding context. In this way\, we control
  the paths of image perception and thereby keep the ambiguity of the render
 ed abstract art to a desirable level. This method is made possible with (1)
  a recent computer vision technique named hierarchical image parsing which 
 decomposes an image into its components\; (2) the computation of a numerica
 l measure of perceptual ambiguity named imperceptibility\, defined as the i
 nformation entropy of different interpretations\, based on a database conta
 ining huge number of annotated image examples\; and (3) our previous work o
 n painterly rendering. Human perceptual experiments show that the rendering
  results have similar ambiguity effects to abstract arts.
SUMMARY:Rendering Abstract Arts by Controlling Imperceptibility
DTSTART:20090428T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:7
SEQUENCE:0
DTEND:20090421T160000
UID:2009-11-17T20:55:58-08:00_146926341@limen.stat.ucla.edu
DESCRIPTION:This seminar\, which will be in question-and-answer format\, wi
 ll survey major developments in the statistical modeling of earthquake occu
 rrences and point process research in general. Topics will include Markov p
 oint process models such as the Stress-Release model\, Hawkes models such a
 s the Epidemic-Type Aftershock Sequence models\, fundamental empirical seis
 mological observations such as Omori's law and the Gutenberg-Richter relati
 on. as well as issues related to the process of writing the book "An Introd
 uction to the Theory of Point Processes"
SUMMARY:Early Days in Point Processes and Statistical Seismology
DTSTART:20090421T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:12
SEQUENCE:0
DTEND:20090310T160000
UID:2009-11-17T20:55:58-08:00_58723126@limen.stat.ucla.edu
DESCRIPTION:In a very real sense\, Google can be thought of as a huge stati
 stical analysis system. This talk will describe how Google uses statistics 
 to turn huge amounts of heterogeneous datainto information about search\, a
 ds\, and advertisers. \n\n Bio: Diane Lambert is a Research Scientist at Go
 ogle\, a fellow of the American Statistical Association and the Institute o
 f Mathematical Statistics.
SUMMARY:Statistics at Google Scale
DTSTART:20090310T150000
DTSTAMP:20091117T205558
LOCATION:8500 Boelter Hall Penthouse
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:47
SEQUENCE:0
DTEND:20071204T160000
UID:2009-11-17T20:55:58-08:00_928457778@limen.stat.ucla.edu
DESCRIPTION:This introductory talk on geostatistics will focus on basic con
 cepts and not the mathematics.  The following topics will be covered: \n\n 
 * Statistical Problem to be solved * Why use any Statistical Approach? * As
 sumptions * Spatial Estimator * Questions to Consider * The Semi-Variogram 
 * Kriging Game * A walk through a geostatistical analysis on a briny Aquife
 r
SUMMARY:Brief Introduction to Geostatistics
DTSTART:20071204T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:55
SEQUENCE:0
DTEND:20071002T160000
UID:2009-11-17T20:55:58-08:00_136369990@limen.stat.ucla.edu
DESCRIPTION:Conditional quantile functions are often estimable under fixed 
 or random censoring without strong distributional assumptions. Applications
  of quantile regression in a wide range of studies have motivated us to tak
 e a hard look at problems arising from censored data. In this talk\, we sha
 ll review the literature on censored quantile regression\, discuss the assu
 mptions underlying various approaches to identifiability and inference\, an
 d describe an efficient estimation method for both right-censored and doubl
 y censored data. We show that most of the existing methods in the literatur
 e are not only imputationally difficult\, but also statistically inferior t
 o a simple mass distribution approach used by Kaplan-Meier. We shall use si
 mple examples to demonstrate the proposed approach\, discuss the practical 
 implications\, and outline some further research problems related to censor
 ed quantile regression. \n\n Biosketch: Dr. Xuming He is Professor of Stati
 stics\, and a Jerry and Ann Nerad Professorial Scholar at the University of
  Illinois at Urbana-Champaign. Dr. He was Director of Illinois Statistics O
 ffice from 2000 to 2003\, overseeing statistical consulting and collaborati
 ve research involving academia\, industry and government agencies.  He was 
 Program Director of Statistics at the National Science Foundation from 2003
  to 2005. He is author/co-author of over 80 scientific articles in the theo
 ry and applications of  robust statistics\,  nonparametric statistical mode
 ling and inference\, and dimension reduction.  He is elected Fellow of the 
 Institute of Mathematical Statistics (IMS) and of the American Statistical 
 Association (ASA)\, and serves on the editorial boards of several premier s
 tatistics journals\, including the Annals of Statistics and JASA.
SUMMARY:Statistical Issues in Censored Regression Quantiles
DTSTART:20071002T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:424
SEQUENCE:0
DTEND:20091201T160000
UID:2009-11-17T20:55:58-08:00_677682196@limen.stat.ucla.edu
DESCRIPTION:In the setting of multivariate analysis\, we discuss Mahalanobi
 s' idea of a distance based on "transforming the coordinates of a statistic
 al field". While Mahalanobis (1936) considered the set of normal distributi
 ons in particular\, we will consider possible extensions to a superset of d
 istributions. The distance can be useful for projection\, as loss function\
 , for creating acceptance regions\, and for hypothesis testing.
SUMMARY:On a Generalized Distance in Statistics &mdash\; Mahalanobis' Idea 
 Revisited
DTSTART:20091201T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:13
SEQUENCE:0
DTEND:20090303T160000
UID:2009-11-17T20:55:58-08:00_319348205@limen.stat.ucla.edu
DESCRIPTION:Copy number variation (CNV) refers to deletion or duplication o
 f a segment of sub-microscopic DNA sequence\, which contributes to a consid
 erable portion of human phenotypic variations and disease susceptibility. I
 n additional to array-CGH\, whole-genome SNP genotyping array provides anot
 her more sensitive method for surveys of CNV. Here\, we attempt to reconstr
 uct CNV from Illumina BeadChip data with two non-parametric total variation
  based methods\, allowing for CNV detection in kilo-base resolution.
SUMMARY:Detection of Copy Number Variants in Human Genome Using Genotype Ar
 ray Data
DTSTART:20090303T150000
DTSTAMP:20091117T205558
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:14
SEQUENCE:0
DTEND:20090224T160000
UID:2009-11-17T20:55:58-08:00_889196721@limen.stat.ucla.edu
DESCRIPTION:MCMC techniques have explored a number of methods for avoiding 
 getting stuck in local minima when sampling the solution space.  However\, 
 it is very difficult for most samplers to jump out of large modes to explor
 e alternative solutions when the distribution being sampled is multi-modal 
 or has multiple solutions.  In this talk I will present an improvement to S
 wendsen-Wang clustering that allows a sampler to make large moves in the so
 lution space by probabilistically forming groups of compatible variable ass
 ignments and swapping them with competing variable assignments.  This allow
 s the sampler to move quickly in the solution space and is ideal for proble
 ms where there may be multiple competing solutions (e.g. an English sentenc
 e may have multiple valid interpretations).  I will present results showing
  that this method out-performs Swendsen-Wang clustering and modern techniqu
 es like Belief Propagation on the Potts and Ising models\, and can also be 
 extended to general labeling problems as well.
SUMMARY:Probabilistic Clustering For Exploring Competing Solutions
DTSTART:20090224T150000
DTSTAMP:20091117T205558
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:58
SEQUENCE:0
DTEND:20070529T160000
UID:2009-11-17T20:55:58-08:00_326583165@limen.stat.ucla.edu
DESCRIPTION:Automatic segmentation and learning of dominant motion patterns
  or activities from a video is an important visual surveillance problem. Mo
 st of the current approaches assume that the observed scene is not crowded\
 , and that reliable tracks of objects are available over longer durations. 
 Therefore\, these approaches are not extendable to more challenging surveil
 lance videos of crowded environments like markets\, subways\, religious fes
 tivals\, parades\, concerts\, football matches etc\, where tracking of indi
 vidual objects is very hard\, if not impossible. \n\n In this talk\, I will
  first a framework for modeling scenes involving high density crowds in whi
 ch Lagrangian particle dynamics are used to segment crowd flows and detect 
 any flow instability. For this purpose flow fields generated by moving crow
 ds are treated as an aperiodic dynamical system which is manifested in term
 s of time dependent optical flow. A grid of particles is overlaid on the fl
 ow field\, and particles are advected using a numerical integration scheme.
  This is followed by the quantification of the amount by which the neighbor
 ing particles have diverged using a Cauchy-Green deformation tensor. The ma
 ximum eigenvalue of this tensor is used to construct a Finite Time Lyapunov
  Exponent (FTLE) field\, which reveals the Lagrangian Coherent Structures (
 LCS) present in the underlying flow. The LCS divides the flow into regions 
 of qualitatively different dynamics and therefore can be used to locate flo
 w segment boundaries. This is done by segmenting the FTLE field using a nor
 malized cuts framework. \n\n Next\, I will present an algorithm for detecti
 ng global motion patterns that exploits the instantaneous motion informatio
 n present in a video instead of long-term motion tracks. A motion pattern i
 s then defined as a group of flow vectors that are part of the same physica
 l process or motion. Algorithmically\, this is accomplished by first detect
 ing the representative modes (sinks) of motion patterns\, followed by the g
 eneration of super tracks\, which coherently represent the discovered motio
 n patterns. \n\n Dr. Mubarak Shah\, Agere Chair Professor of Computer Scien
 ce\, and the founding director of the Computer Visions Lab at the Universit
 y of Central Florida\, is a researcher in a number of computer vision areas
 . Dr. Shah is a fellow of IEEE and IAPR. In 2006\, he was awarded a Pegasus
  Professor award\, the highest award at UCF\, given to a faculty member who
  has made a significant impact on the university\, has made an extraordinar
 y contribution to the university community\, and has demonstrated excellenc
 e in teaching\, research and service. He was an IEEE Distinguished Visitor 
 speaker for 1997-2000 and received IEEE Outstanding Engineering Educator Aw
 ard in 1997. He received the Harris  Corporation's Engineering Achievement 
 Award in 1999\, the TOKTEN awards from UNDP in 1995\, 1997\, and 2000\; Tea
 ching Incentive Program award in 1995 and 2003\, Research Incentive Award i
 n 2003\, Millionaires' Club  awards in 2005 and 2006\, University Distingui
 shed Researcher award in 2007\,  honorable mention for the ICCV 2005 Where 
 Am I? Challenge Problem\, and was nominated for the best paper award in ACM
  Multimedia Conference in 2005.  He is an editor of international book seri
 es on Video Computing\; editor in chief of Machine Vision and Applications 
 journal\, and an associate editor of ACM Computing Surveys journal. He was 
 an associate editor of the IEEE Transactions on PAMI\, and a guest editor o
 f the special issue of International Journal of Computer Vision on Video Co
 mputing.
SUMMARY:Visual Analysis of Crowded Scenes
DTSTART:20070529T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:277
SEQUENCE:0
DTEND:19991012T160000
UID:2009-11-17T20:55:58-08:00_833930715@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Hypertension: 50 Years of Erroneous Thinking\, or How so Many Could
  be so Wrong for so Long
DTSTART:19991012T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:281
SEQUENCE:0
DTEND:19990503T160000
UID:2009-11-17T20:55:58-08:00_139377731@limen.stat.ucla.edu
DESCRIPTION:Both statisticians and biologists have used the one-parameter P
 oisson-Dirichlet distribution since the work of Ewens\, Ferguson\, and King
 man in the 1970's.  Recently\, this family has been extended to include a s
 econd parameter in papers by Pitman and others. \n\n In this talk\, we look
  at both statistical and biological applications of this larger family of d
 istributions.  This includes a larger class of discrete random measures\, a
  new distribution on the probability simplex\, and various results applicab
 le to species sampling.
SUMMARY:Applications of the Two-parameter Poisson-Dirichlet Distribution
DTSTART:19990503T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:18
SEQUENCE:0
DTEND:20090127T160000
UID:2009-11-17T20:55:58-08:00_689834404@limen.stat.ucla.edu
DESCRIPTION:The technology to harvest electricity from wind energy is now a
 dvanced enough to make entire cities powered by it a reality. High-quality 
 short-term forecasts of wind speed are vital to making this a more reliable
  energy source. Gneiting et al. (2006) have introduced a model for the aver
 age wind speed two hours ahead based on both spatial and temporal informati
 on. The forecasts produced by this model are accurate\, and subject to accu
 racy\, the predictive distribution is sharp\, i.e.\, highly concentrated ar
 ound its center. However\, this model is split into nonunique regimes based
  on the wind direction at an off-site location. This paper both generalizes
  and improves upon this model by treating wind direction as a circular vari
 able and including it in the model. It is robust in many experiments\, such
  as predicting at new locations. We compare this with the more common appro
 ach of modeling wind speeds and directions in the Cartesian space and use a
  skew-t distribution for the errors. The quality of the predictions from al
 l of these models can be more realistically assessed with a loss measure th
 at depends upon the power curve relating wind speed to power output. This p
 roposed loss measure yields more insight into the true value of each model’
 s predictions.
SUMMARY:Powering Up with Space-Time Wind Forecasting
DTSTART:20090127T150000
DTSTAMP:20091117T205558
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:419
SEQUENCE:0
DTEND:20091020T160000
UID:2009-11-17T20:55:58-08:00_53489717@limen.stat.ucla.edu
DESCRIPTION:Classification is a fundamental task in statistical machine lea
 rning\, with numerous practical applications in many fields. Classification
  using boosting (and trees) is particularly popular in industry. This talk 
 will present the general idea of ABC-Boost (adaptive base class boosting) f
 or multi-class classification and its two specific implementations named AB
 C-MART and ABC-LogitBoost. \n\n The original MART (Friedman 2001) and Logit
 Boost (Friedman\, Hastie\, and Tibshirani\, 2000) algorithms are highly inf
 luential in the field of statistical machine learning. We show that\, on ma
 ny public datasets\, ABC-MART could improve MART roughly by 10% (relatively
 ) in terms of the mis-classification errors. Furthermore\, ABC-Logitboost c
 ould improve ABC-MART roughly by another 10%. \n\n Bio:   Ping Li is an ass
 istant professor in the Department of  Statistical Science at Cornell Unive
 rsity. In 2007\, he graduated from Stanford University\, with a Ph.D. in St
 atistics\, an M.S. in Computer Science\, and an M.S. in Electrical Engineer
 ing. His research interests include (1) fundamental randomized algorithms f
 or processing massive (and possibly streaming) datasets\; (2) statistical m
 achine learning. Ping Li is a recipient of the ONR (Office of Naval Researc
 h) Young Investigator Award in 2009.
SUMMARY:Adaptive Base Class Boosting (ABC-Boost) for Multi-Class Classifica
 tion
DTSTART:20091020T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:19
SEQUENCE:0
DTEND:20090120T160000
UID:2009-11-17T20:55:58-08:00_152589244@limen.stat.ucla.edu
DESCRIPTION:Recursive Compositional Models (RCMs) are used for modeling the
  visual patterns of images and objects. The key design principle is recursi
 ve compositionality.  Visual patterns are represented by RCMs in a hierarch
 ical form where complex structures are composed of more elementary structur
 es. Probabilities are defined over these structures exploiting properties o
 f the hierarchy (e.g. long range spatial relationships can be represented b
 y local potentials). The compositional nature of this representation enable
 s efficient learning and inference algorithms. Hence the overall architectu
 re of RCMs provides a balance between statistical and computational complex
 ity. \n\n Work with Long Zhu.
SUMMARY:Recursive Compositional Models
DTSTART:20090120T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:177
SEQUENCE:0
DTEND:20030527T160000
UID:2009-11-17T20:55:58-08:00_170840600@limen.stat.ucla.edu
DESCRIPTION:In joint work with Silke Rolles\, Bayesian methods have been de
 veloped for analyzing Markov Chain data when the chain is reversible.  This
  includes closed form conjugate priors\, a version of De Finetti's Theorem 
 and characterization theorems.  Examples are given.
SUMMARY:Bayesian Analysis of Reversible Markov Chains
DTSTART:20030527T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:178
SEQUENCE:0
DTEND:20030530T160000
UID:2009-11-17T20:55:58-08:00_843506654@limen.stat.ucla.edu
DESCRIPTION:In this talk we shall first give a brief overview of financial 
 engineering\, and then discuss various stochastic models\, including a stoc
 hastic model for biotech and internet stocks\, a jump diffusion model for o
 ption pricing\, and pricing of path-dependent options by using the classica
 l renewal theory. The title of the talk indicates that our goal is to prese
 nt the results in such a way that it may be not only of interest to mathema
 tical finance people but also to the general audience of both statistical a
 nd financial communities\; in particular\, we promise that there will be no
  mathematical formula in the talk. By the way\, if you want to understand t
 he title of the talk\, please come to the seminar.
SUMMARY:Financial Engineering in the Wonderland: An Adventure with Riemann\
 , Lisa\, and Dickens
DTSTART:20030530T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:179
SEQUENCE:0
DTEND:20030513T160000
UID:2009-11-17T20:55:58-08:00_281447735@limen.stat.ucla.edu
DESCRIPTION:Estimating and visualizing high-order statistics of multivariat
 e data is important for analysis\, synthesis and visualization in science a
 nd engineering. Often\, data consists of measurements on an underlying doma
 in\, such as space or time. Examples include images\, audio signals and tex
 t\, where the domains are 2-D space\, 1-D time and 1-D symbol index. We int
 roduce a model called the 'epitome' that can simultaneously represent multi
 -scale high-order statistics as a set of parameters on the same domain as t
 he input data. A cost function measures how well multi-scale patches drawn 
 from the input data match the epitome and this cost function can be optimiz
 ed efficiently using the EM algorithm. Our technique reduces a large number
  of high-order statistics to an intuitive\, compact representation that is 
 suitable for a variety of data processing applications. We demonstrate our 
 method using problems of object detection\, texture segmentation and image 
 retrieval. \n\n Joint work with Nebojsa Jojic\, Microsoft Research\, and An
 itha Kannan\, University of Toronto. \n\n Bio: Brendan J. Frey was born on 
 August 29\, 1968\, in Calgary\, Alberta near the foothills of the Rocky Mou
 ntains\, where he enjoyed hiking and camping with his family. In 1979\, he 
 started writing computer programs\, attaching sensors to his home computer\
 , and building simple robots. His first publication (Run Magazine\, 1981) d
 escribes software for simulating a generative model of images\, where the p
 ixel intensities are independent. His academic education was in the areas o
 f physics\, engineering and computer science\, culminating with a doctorate
  from Geoffrey Hinton's Neural Networks Research Group at the University of
  Toronto. From 1997 to 1999\, Frey was a Beckman Fellow at the University o
 f Illinois at Urbana-Champaign\, where he continues to be an adjunct facult
 y member in Electrical and Computer Engineering. From 1998 to 2001\, he was
  a faculty member in Computer Science at the University of Waterloo. Curren
 tly\, Frey is head of the Probabilistic and Statistical Inference Group\, i
 n the Department of Electrical and Computer Engineering at the University o
 f Toronto\, and consults for Microsoft Research Redmond. He has received se
 veral awards\, given over 40 invited talks and published over 100 papers on
  inference and estimation in complex probability models for machine learnin
 g\, computer vision\, speech processing\, image processing\, and iterative 
 decoding. In 2003\, he co-chaired the Canadian Workshop on Information Theo
 ry and also the Workshop on Artificial Intelligence and Statistics. He acte
 d as Co-Editor-in-Chief of the February 2000 special issue of the IEEE Tran
 sactions on Information Theory\, titled Codes on Graphs and Iterative Algor
 ithms\, and is currently an Associate Editor of the IEEE Transactions on Pa
 ttern Analysis and Machine Intelligence.
SUMMARY:Learning the 'Epitome' of an Image
DTSTART:20030513T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:21
SEQUENCE:0
DTEND:20090106T160000
UID:2009-11-17T20:55:58-08:00_907490911@limen.stat.ucla.edu
DESCRIPTION:This talk is an overview of (social) network modeling from the 
 perspective of a statistician. \n\n Network models are widely used to repre
 sent relational information among interacting units and the implications of
  these relations.  In studies of social networks recent emphasis has been p
 laced on random graph models where the nodes usually represent individual s
 ocial actors and the edges represent a specified relationship between the a
 ctors. \n\n The modeling of social networks is\, and has been\, broadly mul
 tidisciplinary with significant contributions from the social\, natural and
  mathematical sciences.  This has lead to a plethora of terminology\, and n
 etwork conceptualizations commensurate with the varied objectives of networ
 k analysis.  As a primary focus of the social sciences has been the represe
 ntation of social relations with the objective of understanding social stru
 cture\, social scientists have been central to this development. \n\n Expon
 ential family random graph models attempt to represent the complex dependen
 cies in networks in a parsimonious\, tractable and interpretable way.  A ma
 jor barrier to the application of such models has been lack of understandin
 g of model behavior and a sound statistical theory to evaluate model fit.  
 This problem has at least three aspects: the specification of realistic mod
 els\; the algorithmic difficulties of the inferential methods\; and the ass
 essment of the degree to which the network structure produced by the models
  matches that of the data. \n\n In this talk we review progress that has be
 en made on networks observed in cross-sectional or longitudinally.  We cons
 ider issues of the sampling of networks and partially-observed networks.  W
 e also consider latent cluster random effects models. \n\n We illustrate th
 ese methods using the "statnet" open-source software suite (http://statnet.
 org). \n\n Biographical: Mark S. Handcock is Professor and Chair of Statist
 ics\, Department of Statistics\, University of Washington\, Seattle.  His w
 ork is based largely on motivation from questions in the social sciences.  
 Recent focus has been on the development of statistical models for the anal
 ysis of social network data\, spatial processes and demography.  He receive
 d his B.Sc.  from the University of Western Australia and his Ph.D.  from t
 he University of Chicago.  Descriptions of his work are available at http:/
 /www.stat.washington.edu/handcock.
SUMMARY:Statistical Modeling of Social Networks
DTSTART:20090106T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:22
SEQUENCE:0
DTEND:20081202T160000
UID:2009-11-17T20:55:58-08:00_586237931@limen.stat.ucla.edu
DESCRIPTION:The count data we encounter in practice often have many zero-va
 lued observations. Analyzing such data without any consideration for excess
  zeros could produce misleading results. We first show possible consequence
 s of ignoring excess zeros in analysis. We then show some results of an ana
 lysis of shark bycatch data collected in a tuna purse-seine fishery. Finall
 y\, we introduce our current research\, a new feature extraction method for
  very non-normal data such as count data with many zeros.
SUMMARY:Statistical Challenges for Modeling Data with Many Zero-valued Obse
 rvations
DTSTART:20081202T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:59
SEQUENCE:0
DTEND:20070522T160000
UID:2009-11-17T20:55:58-08:00_74245949@limen.stat.ucla.edu
DESCRIPTION:Data obtained as counts over geographic regions can be describe
 d as Poisson random variables\, and this work models the Poisson expected v
 alues.  There are two major points here.  (1) The model parameterizes hot s
 pots in a way that allows the estimation of the hot spot coordinates.  (2) 
 Spatial integrals are computed through Green's theorem.  Two disease data p
 roblems are used as examples.
SUMMARY:Hot Spot Spatial Models for Data Reported as Counts over Geographic
  Areas
DTSTART:20070522T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:24
SEQUENCE:0
DTEND:20081104T160000
UID:2009-11-17T20:55:58-08:00_919464032@limen.stat.ucla.edu
DESCRIPTION:In this talk\, I will present a few Markov chain examples for w
 hich knowledge of multivariate orthogonal polynomials turns out helpful. \n
 \n 1.  How many bulbs light up?  There are N light bulbs which are initiall
 y all off. On the first day\, we randomly choose one bulb and switch it on.
  On the second day\, we randomly choose two bulbs. If a chosen bulb is on\,
  we switch it off\; if off\, we switch it on. In general\, on day t\, we ra
 ndomly choose t bulbs and press their switches. The question is how many bu
 lbs are on at end of day $N$. This model\, initially studied by Rao\, Rao\,
  and Zhang (2006)\, is motivated by a dermal patch problem in medicine. We 
 give explicit expressions for the mean\, variance and distribution of the n
 umber of on bulbs at any time. \n\n 2. How soon are friends meeting? Geogra
 phically\, Sequoia Hall\, Gates Hall and Bytes Cafe form a triangle. At beg
 inning\, three persons are at distinct places\, say Abel in Sequoia Hall\, 
 Betty in Gates Hall\, and Cal in Bytes Cafe. At each step\, one of them (ra
 ndomly chosen) performs one step of random walk\, i.e.\, the chosen individ
 ual flips a coin and walks to one of neighboring buildings. One question is
  how long it takes for all three persons to meet at Bytes Cafe. The answer 
 is on average it takes 31 steps and the whole distribution of the coalescen
 t time can also be obtained. This is one example of the coalescent times of
  multi-person random walk on graph studied by Tian and Liu (2007). We provi
 de a more general solution to this problem. \n\n 3. How fast does a populat
 ion mix? Moran's model is one of the most important stochastic processes in
  population genetics. Recently\, similar models revived in community ecolog
 y under the so-called Unified Neutral Theory of Biodiversity and Biogeograp
 hy (UNTB) proposed by Hubbell (2001). When testing the neutral theory\, eit
 her in population genetics or community ecology\, biologists want to know w
 hether the population is at equilibrium or not. We give results for the con
 vergence rates of Moran type processes in chi-square and separation distanc
 es. \n\n 4.  Convergence rate of an algorithm from Bayesian image analysis.
  In Bayesian image analysis\, often people put a multivariate Gaussian prio
 r on the image and assume that the observed image is the original image cor
 rupted by Gaussian noise. Gibbs samplers are used to sample from the poster
 ior distribution\, which is again a multivariate Gaussian. We give converge
 nce rate of one version of such Gibbs samplers. \n\n The first two examples
  are from joint work with Kenneth Lange. The last two examples are from joi
 nt work with Kshitij Khare.
SUMMARY:Examples of Markov Chains with Multivariate Orthogonal Polynomial E
 igenfunctions
DTSTART:20081104T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:25
SEQUENCE:0
DTEND:20081028T160000
UID:2009-11-17T20:55:58-08:00_177520434@limen.stat.ucla.edu
DESCRIPTION:Cluster analysis focuses on grouping observations into classes 
 according to an appropriate measure of proximity. Many questions about clus
 tering in a Euclidean n-space can be answered using the fact that the neare
 st-neighbor distance d for a homogeneous Poisson field has the Weibull dist
 ribution with the shape parameter n. The situation becomes more complicated
  for processes that evolve in space-time and whose events can differ by siz
 e. The main question is how to measure the distance between events in such 
 a process\, so that all space-time-size characteristics are taken into acco
 unt. We propose a general framework for cluster analysis of marked point pr
 ocesses in time-space-size domain. Specifically\, we define an asymmetric p
 seudo-distance n(t\,x\,m) for a point process N(t\,x) with time component t
 \, space component x\, and size mark m and show that under some natural ass
 umptions this distance has the Weibull distribution. Furthermore\, we study
  the joint 2D distribution of size-rescaled time and space components of th
 e distance n and show how to use it for cluster analysis and inference. Fin
 ally\, we apply the proposed analyzes to solve the problem of aftershock id
 entification in the observed seismicity of California. \n\n Aftershock iden
 tification problem. The centennial observations of the world seismicity hav
 e revealed a wide variety of clustering phenomena that unfold in the space-
 time-energy domain and provide the most reliable information about earthqua
 ke dynamics. Nevertheless\, there is neither a unifying theory nor a conven
 ient statistical apparatus that would naturally account for the different t
 ypes of seismic clustering. Notably\, there is no objective (i.e.\, based o
 n statistically significant separation) method for identifying earthquake a
 ftershocks\, the most prominent part of seismic clustering. We use the prop
 osed cluster analysis to identify the earthquake clustered and homogeneous 
 parts using no apriori parameters and show that the two parts are statistic
 ally different\, and that the clustered part mainly consists of events comm
 only considered to be aftershocks. This finding serves as a basis for an ob
 jective non-parametric aftershock identification procedure. \n\n Bio-sketch
  Ilya Zaliapin is an Assistant Professor in the Department of Mathematics a
 nd Statistics at University of Nevada\, Reno\, which he has joined in 2006.
  Dr. Zaliapin's research is focused on environmental statistics and geo-haz
 ards assessment and forecasting. In particular\, he works on hierarchical a
 ggregation processes\, heavy-tailed models\, and multiscale analysis of tim
 e series with applications to seismology\, hydrology\, climate studies\, an
 d cell biology. Ilya Zaliapin got his PhD in Physics and Mathematics (Proba
 bility and Statistics) from the Russian Academy of Sciences in 1999\; befor
 e joining UNR\, he was a postdoc (1999-2001) and then a researcher (2001-10
 06) in the Institute of Geophysics and Planetary Physics at UCLA.
SUMMARY:Cluster Analysis for Marked Point Processes  with Application to th
 e Aftershock Detection Problem
DTSTART:20081028T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:71
SEQUENCE:0
DTEND:20070130T160000
UID:2009-11-17T20:55:58-08:00_96327928@limen.stat.ucla.edu
DESCRIPTION:In this talk\, I'd like to explore some connections between sta
 tistics and abstract arts. More specifically\, we argue that the creation o
 f abstract arts might possibly be formulated in  statistical modeling and s
 tochastic simulation. As a results\, computers can generate some abstract a
 rts through random sampling. Actually\, this shouldn't be too surprising\, 
 given that computers can now render realistic pictures and cartoon animatio
 ns. \n\n This talk is inspired by a recent talk of Jean-Michel Morel\, an a
 pplied mathematician at Ecole Normale Superieure de Cachan. Many slides are
  directly borrowed from Dr. Morel.
SUMMARY:Can We Formulate Abstract Arts in Statistics?
DTSTART:20070130T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:381
SEQUENCE:0
DTEND:19970303T160000
UID:2009-11-17T20:55:58-08:00_201288013@limen.stat.ucla.edu
DESCRIPTION:Fitting a circle to a set of data points on a plane is very com
 mon in engineering and science. Important practical problems include the ch
 oice of locations of measurement on a circular feature\, and the estimation
  of its center and radius after taking sample points. Both issues have not 
 been studied in depth in the statistical literature. For the design problem
 \, only some simulation results are available.  Even for some simple circul
 ar models\, consistent estimates have not yet been rigorously established i
 n the literature.  Consequently\, estimation methods for the center and rad
 ius of a circular feature commonly used in practice are far from satisfacto
 ry. \n\n In this talk\, first\, under Berman's (1983) bivariate four-parame
 ter model\, <span class='math'>\Phi</span>-optimality is defined and shown 
 to be equivalent to all <span class='math'>\phi_{p}</span>-criteria with <s
 pan class='math'>p \in [-\infty\, 1)</span>\, which include the well-known 
 <span class='math'>A</span>-\, <span class='math'>D</span>-\, and <span cla
 ss='math'>E</span>-criteria.  <span class='math'>\Phi</span>-optimal exact 
 designs on a circle or a circular arc are then analytically derived for any
  sample size and sampling range. These results provide guidelines for users
  on sampling method and sample size selection. \n\n For the estimation prob
 lem\, asymptotic results for general estimating equations are established. 
 These results are of intrinsic interest\, and also lead to some classical a
 symptotic results of maximum likelihood estimation.  They are used to estab
 lish consistent and asymptotically normal estimators for circular functiona
 l models. The inconsistency of least squares estimators is also obtained.
SUMMARY:Design and Estimation in Circular Measurement Error Models
DTSTART:19970303T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:27
SEQUENCE:0
DTEND:20081007T160000
UID:2009-11-17T20:55:58-08:00_756126442@limen.stat.ucla.edu
DESCRIPTION:I will review concepts\, principles\, and mathematical tools th
 at were found useful in applications involving causal inference.  The princ
 iples are based on structural-model semantics\, in which functional (or cou
 nterfactual) relations represent physical processes. This semantical framew
 ork\, enriched with a few ideas from logic and graph theory\, gives rise to
  a complete\, coherent\, and friendly calculus of causation that unifies th
 e structural\, graphical and potential-outcome approaches to causation and 
 resolves long-standing problems in several of the sciences.  These include 
 questions of confounding\, causal effect estimation\, policy analysis\, leg
 al responsibility\, effect decomposition\, instrumental variables\, and the
  integration of data from diverse studies. \n\n Reference: J. Pearl\, Causa
 lity (Cambridge University Press\, 2000) (http://bayes.cs.ucla.edu/jp_home.
 html) \n\n Tutorials: http://bayes.cs.ucla.edu/IJCAI99/ ftp://ftp.cs.ucla.e
 du/pub/stat_ser/R271.pdf ftp://ftp.cs.ucla.edu/pub/stat_ser/R273.pdf \n\n B
 io Judea Pearl Judea Pearl is a professor of computer science and statistic
 s at the University of California\, Los Angeles. He joined the faculty of U
 CLA in 1970\, where he currently directs the Cognitive Systems Laboratory a
 nd conducts research in artificial intelligence\, human reasoning and philo
 sophy of science.  He has authored three books: Heuristics (1984)\, Probabi
 listic Reasoning (1988)\, and Causality (2000).  A member of the National A
 cademy of Engineering\, and a Founding Fellow the American Association for 
 Artificial Intelligence (AAAI)\, Judea Pearl is the recipient of the IJCAI 
 Research Excellence Award for 1999\, the London School of Economics Lakatos
  Award for 2001\, the ACM Alan Newell Award for 2004\, and the 2008 Benjami
 n Franklin Medal of Computer and Cognitive Science from the Franklin Instit
 ute.
SUMMARY:The Mathematics of Causal Inference
DTSTART:20081007T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:109
SEQUENCE:0
DTEND:20051213T160000
UID:2009-11-17T20:55:58-08:00_24308171@limen.stat.ucla.edu
DESCRIPTION:Sensitivity of Earth's climate to increasing amounts of atmosph
 eric carbon dioxide is a topic of general and scientific interest\, and pub
 lic policy.  Today's global climate models predict that the strongest depen
 dences of surface temperatures on increasing atmospheric carbon dioxide lev
 els will occur in the Arctic.  Ascertaining the properties of clouds in the
  Arctic via conventional satelliate images is a challenging problem because
  liquid and ice water cloud particles often have similar properties to the 
 snow and ice particles that compose snow- and ice-covered surfaces. Without
  accurate characterization of clouds over the Arctic we will not be able to
  assess the impact of clouds on the flow of solar and terrestrial electroma
 gnetic radiation through the Arctic atmosphere and we will not be able to a
 scertain whether they are changing in ways that enhance or ameliorate futur
 e warming in the Arctic. With the launch of the Multi-angle Imaging Spectro
 Radiometer (MISR) and Moderate Resolution Imaging Spectrometer (MODIS) by N
 ASA in 1999\, novel electromagnetic radiation measurements made at many ang
 les (from MISR) and across many narrow wavelength (hyperspectral) regions i
 n the visible and infrared (by MODIS) became available for scientific study
 . \n\n In this talk\, we will report on an on-going collaborative effort to
 wards fast and effective arctic cloud detection algorithms based on both MI
 SR and MODIS measurements. After much interaction with the MISR team at JPL
 \, we arrived at three physically meaningful features from MISR for our clu
 stering  algorithm\, Enhanced Linear Correlation Matching Clustering (ELCMC
 ). This algorithm thresholds the three features with two fixed thresholds a
 nd one adaptive threshold which is found by an EM algorithm. It is robust a
 nd computationally fast for on-line processing of MISR images.  Further imp
 rovements over ELCMC could be achieved by using concensus lables of ELCMC-M
 ISR and MODIS to train Fisher's Quadratic Discriminative Analysis (QDA). Su
 pport vector machines (SVMs) can obtain better accuracies when compared wit
 h expert labels for some images than QDA\, but they are much slower than QD
 A. Binning is then applied to the features before SVMs to speed up the SVM 
 computation substantially with comparable cloud detection results.  The clo
 ud labels from different algorithms are compared with the best "ground trut
 h" available in large quantities -- the expert labels and show extremely go
 od agreements. Moreover\, our ELCMC-QDA algorithm outputs classification pr
 obabilities around 0.5 for pixels where the expert is not comfortable to pr
 ovide labels. This further confirms the agreement between our algorithm and
  the expert opinion. \n\n (This talk is based on joint work with Tao Shi at
  Ohio State University\, Eugene Clothiaux at Penn State University\, and Am
 y Braverman at NASA's Jet Propulsion Lab.)
SUMMARY:Arctic Cloud Detection using Multi-Angle and Hyperspectral Satellit
 e Images
DTSTART:20051213T150000
DTSTAMP:20091117T205558
LOCATION:6629 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:28
SEQUENCE:0
DTEND:20080930T160000
UID:2009-11-17T20:55:58-08:00_490156256@limen.stat.ucla.edu
DESCRIPTION:Neuroimaging is a field in which statistics is a key tool for d
 iscovering how our brains operate.  Currently there do not exist laboratory
  tests for diagnose of Schizophrenia and Alzheimer's disease\, yet machine 
 learning methods are able to accomplish this without any knowledge of the p
 athology behind these disorders.  In this talk\, I will present a pattern r
 ecognition method for classification and discrimination of fMRI Brain scans
 .  This method will classify based on the native patterns taken by the brai
 n over time and will not require spatial alignment of scans across subjects
 .  The new method will be contrasted with existing methods in this field.
SUMMARY:Mind Reading:  Computer Classification of Schizophrenic and Alzheim
 er's Disease fMRI Scans using Machine Learning
DTSTART:20080930T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:29
SEQUENCE:0
DTEND:20080603T160000
UID:2009-11-17T20:55:58-08:00_542061045@limen.stat.ucla.edu
DESCRIPTION:I'll begin by discussing Sage (http://www.sagemath.org)\, an op
 en-source math software project. Sage includes a number of tools of interes
 t to statisticians\, including R\, rpy\, and Numpy/Scipy\, and I'll try to 
 explain what Sage has to offer beyond just including these packages. \n\n I
 n the second part of my talk\, I'll quickly introduce some recent results a
 nd conjectures in number theory (such as the Sato-Tate Conjecture)\, and di
 scuss some of the interesting statistical questions that arise in trying to
  understand and extend these results.
SUMMARY:Sage and Statistics
DTSTART:20080603T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:30
SEQUENCE:0
DTEND:20080520T160000
UID:2009-11-17T20:55:58-08:00_460295393@limen.stat.ucla.edu
DESCRIPTION:Structural equation models (SEM) is a statistical methodology w
 idely used in behavioral and social sciences\, economics\, marketing\, and 
 other disciplines\, to analyze multivariate observational data.  The genera
 lity of SEM stems from the fact that it allows for measurement error\, simu
 ltaneous equations\, and multiple-indicators of latent (theoretical) variab
 les of the model\, single and multple-group data\, as well as alternative e
 stimation methods.  The factor analysis model developed in psychometrics an
 d the simultaneous equation models of econometrics are classical examples o
 f SEM.  User friendly software such as LISREL\, EQS\, Mplus\, and others\, 
 have contributed to the popularity of SEM among applied researchers in many
  disciplines.  Model specification is of key importance in SEM\, since typi
 cally a model represents a particular view (theory) of the interrelationshi
 p among variables (latent and observed).  Validity of the model is usually 
 assessed by means of a chi-square goodness-of-fit test\, with the power of 
 the test being then of relevance for drawing  appropriate conclusions from 
 the test. \n\n In this talk\, we re-visit goodness-of-fit testing in struct
 ural equation models and a procedure to approximate the power of the test. 
  We investigate the variation of the power of the chi-square goodness-of-fi
 t under different conditions of non-normality of the data.  It is shown tha
 t under certain conditions of the model\, identified as conditions for asym
 ptotic robustness (AR)\, power is a function only of deviations from the mo
 ment structure implied by the model\, with power being insensitive to devia
 tions from non-normality of the latent components of the model.  When AR do
 es not hold\, however\, power is affected by the specific non-normality of 
 the random components of the model.  We also show that under deviation from
  AR\, a scaled chi-square goodness-of-fit permits proper model testing and 
 power analysis in a simple form.  A factor analysis model context and simul
 ated data is used as an illustration.
SUMMARY:Power Analysis in Structural Equation Models with Non-normal Data
DTSTART:20080520T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:31
SEQUENCE:0
DTEND:20080513T160000
UID:2009-11-17T20:55:58-08:00_521918637@limen.stat.ucla.edu
DESCRIPTION:Stochastic models for marked point processes and their characte
 ristics (like Ripley's K-function and the pair correlation function) are co
 nsidered for statistical analysis of labelling patterns of mammary carcinom
 a cell nuclei on histological sections. \n\n In diagnostic histopathology\,
  it is a routine application to use immunohistochemical stains to study the
  labelled  pattern of tumour cell nuclei. Using such methods\, it is possib
 le to classify observed tumour cell nucleus profiles on histological sectio
 ns in a binary manner into two categories\, i.e. stained vs.  unstained nuc
 lei (labelled vs. unlabelled\, positive vs. negative nuclei). In the domain
  of mammary cancer\, a relevant immunohistochemical marker is the MIB-1 sta
 in\, which decorates specifically the nuclei of proliferating (dividing) ce
 lls\, whereas the nuclei of nondividing cells are negative. \n\n It is of s
 cientific interest to find out whether the labelled nuclei are compatible w
 ith a random thinning of all nuclei. In this case\, they would be generated
  from the mother process of all nuclei simply by random labelling. The redu
 ced second moment functions K(r) of the labelled and the unlabelled points 
 would then be identical.  The same would hold for the pair correlation func
 tions g(r). The alternative hypothesis is\, that the second-order propertie
 s of the processes of the labelled and unlabelled points are systematically
  different. \n\n Twenty cases of invasive mammary ductal carcinomas were st
 udied.  The planar coordinates of the tumor cell nuclei from two rectangula
 r visual fields per case were recorded (384-1387 points per field\, of whic
 h 3-27 % were labelled). Subsequently\, for each visual field the following
  investigations were performed by using GeoStoch\, a Java-based open-librar
 y system\, and with the software package  SpatStat. \n\n - Estimation of th
 e explorative summary characteristics K(r) and g(r) \n\n - Fitting of the p
 arameters of a stationary Strauss hard core model to the observed point pat
 terns \n\n - Estimation of two distance-dependent Simpson indices \n\n - Mo
 nte Carlo tests on the null hypothesis of random labelling \n\n Significant
  differences between the mean K-functions and the mean g-functions of the l
 abelled and the unlabelled nuclei were found.  Moreover\, the mean interact
 ion parameter of the stationary Strauss hard core model was significantly h
 igher for the labelled nuclei than for the unlabelled nuclei. The estimates
  of the two distance-dependent Simpson indices showed a tendency of the poi
 nts towards a positive spatial correlation. In the Monte Carlo tests\, the 
 null hypothesis of random labelling was rejected for the majority of the vi
 sual fields. \n\n These four lines of investigation led to the concordant c
 onclusion\, that the labelling of mammary carcinoma nuclei by MIB-1 does no
 t simply result from a random labelling of the nuclei. The data suggest tha
 t the second-order properties of the point process of the labelled nuclei a
 re significantly different from those of the unlabelled nuclei. The process
  of the labelled nuclei shows a higher degree of clustering (increased stre
 ngth of interaction) than the process of the unlabelled points. \n\n The ta
 lk is based on joint research with Stefanie Eckel\, Frank Fleischer and Tor
 sten Mattfeldt.
SUMMARY:Statistical Analysis of Labelling Patterns of Mammary Carcinoma Cel
 l Nuclei on Histological Sections
DTSTART:20080513T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:33
SEQUENCE:0
DTEND:20080429T160000
UID:2009-11-17T20:55:58-08:00_596711389@limen.stat.ucla.edu
DESCRIPTION:I will describe a method for organizing multiple visual categor
 ies into a taxonomy. Such organization becomes crucial as the number of ava
 ilable categories increases. The proposed method extends current non-parame
 tric Bayesian techniques such as Nested Chinese Restaurant Process (NCRP). 
 The method is completely unsupervised. It discovers commonalities among ima
 ges and exploits these commonalities to represent images compactly in a hie
 rarchical manner. Visual categories emerge and become organized in a taxono
 my automatically during this process.
SUMMARY:Unsupervised Learning of Visual Taxonomies
DTSTART:20080429T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:35
SEQUENCE:0
DTEND:20080415T160000
UID:2009-11-17T20:55:58-08:00_90135355@limen.stat.ucla.edu
DESCRIPTION:Automatic localization\, tracking\, and event detection in vide
 os of crowded environments is an important visual surveillance problem. Des
 pite the sophistication of current surveillance systems\, they have not yet
  attained the desirable level of applicability and robustness required for 
 handling crowded scenes like parades\, concerts\, football matches\, train 
 stations\, airports\, city centers\, malls etc. \n\n In this talk\, I will 
 first present a framework for segmenting scenes into dynamically distinct c
 rowd regions using Lagrangian particle dynamics. For this purpose\, the spa
 tial extent of the video is treated as a phase space of a non-autonomous dy
 namical system where transport from one region of the phase space to the ot
 her is controlled by the optical flow. A grid of particles is advected thro
 ugh the phase space using the optical flow using a numerical integration sc
 heme\, and the amount by which neighboring particles diverge is quantified 
 by using a Cauchy-Green deformation tensor. The maximum eigenvalue of this 
 tensor is used to construct a Finite Time Lyapunov Exponent (FTLE) field\, 
 which reveals the time-dependent invariant manifolds of the non-autonomous 
 dynamical system which are called Lagrangian Coherent Structures (LCS). The
  LCS in turn divides the crowd flow into regions of different dynamics\, an
 d\, therefore\, can be used to the segment the scene into distinct crowd re
 gions. The segmentation is then employed to detect any change in the behavi
 or of the crowd over time. Next\, I will present an algorithm for tracking 
 individual targets in high density (hundreds of people) crowded scenes. The
  novelty of the algorithm lies in a scene structure based force model\, whi
 ch is used in conjunction with the available appearance information for tra
 cking individuals in a complex crowded scene. The key ingredients of the sc
 ene structure force model are three fields namely\, 'Static Floor Field' (S
 FF)\, 'Dynamic Floor Field' (DFF)\, and 'Boundary Floor Field' (BFF). These
  fields determine the probability of a person moving from one location to a
 nother in a way that the object movement is more likely in the direction of
  higher fields.
SUMMARY:Segmentation and Tracking in Crowded Scenes
DTSTART:20080415T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:36
SEQUENCE:0
DTEND:20080408T160000
UID:2009-11-17T20:55:58-08:00_78191459@limen.stat.ucla.edu
DESCRIPTION:Probabilistic weather forecasting consists of finding joint pre
 dictive probability distributions of future weather quantities or events\, 
 which is critical for weather-related decision-making.  It is typically don
 e by using a numerical weather prediction model\, perturbing the inputs to 
 the model (initial conditions and physics parameters)\, and running the mod
 el forward for each perturbed set of inputs.  The result is then viewed as 
 an ensemble of forecasts and often interpreted as a sample from the joint p
 redictive probability distribution.  However\, forecast ensembles typically
  are underdispersed and subject to biases\, so statistical postprocessing i
 s required to obtain calibrated probabilistic forecasts.  I will review rec
 ent joint work in this area\, focusing on the Bayesian model averaging (BMA
 ) approach to temperature and precipitation forecasting. These methods have
  been applied to the University of Washington mesoscale ensemble over the P
 acific Northwest\, and postprocessed probabilistic BMA forecasts are availa
 ble in real time at http://probcast.washington.edu.
SUMMARY:Probabilistic Weather Forecasting
DTSTART:20080408T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:46
SEQUENCE:0
DTEND:20071211T160000
UID:2009-11-17T20:55:58-08:00_658096282@limen.stat.ucla.edu
DESCRIPTION:The analysis of time series of complex data such as intervals a
 nd histograms is providing further insights in the symbolic data analysis (
 SDA) field. In this talk\, we shall review the literature on symbolic data 
 and show some forecasting methods. \n\n We shall use simple examples\, most
  of them from financial markets\, to demonstrate the methods developed\, di
 scuss the practical implications\, and outline some further research proble
 ms related to forecasting symbolic data as the Bayesian approach.
SUMMARY:Forecasting Methods for Symbolic Data — Some Comments for Bayesian 
 Developments
DTSTART:20071211T150000
DTSTAMP:20091117T205558
LOCATION:5200 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:38
SEQUENCE:0
DTEND:20080311T160000
UID:2009-11-17T20:55:58-08:00_199766848@limen.stat.ucla.edu
DESCRIPTION:Much computer vision research can be described as pattern recog
 nition on separate subcomponents\, such as segmentation\, object recognitio
 n\, image categorization\, etc.   Such work is essential\, but to truly und
 erstand the scene we must look beyond the image plane and reason about obje
 cts and surfaces in the world.   This presents two major challenges: 1) How
  can we get any 3D scene properties from a single 2d projection?\; and 2) H
 ow can we robustly interrelate thousands of visual concepts?  I will descri
 be my research that begins to answer each of these questions.  The key is t
 o learn statistical models of the visual world while avoiding brittle assum
 ptions and early decisions.   This work provides heartening evidence of pro
 gress in scene understanding\, but much more remains to be done.
SUMMARY:Seeing the World Behind the Image
DTSTART:20080311T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:40
SEQUENCE:0
DTEND:20080219T160000
UID:2009-11-17T20:55:58-08:00_510003611@limen.stat.ucla.edu
DESCRIPTION:In the past decade\, we have witnessed the revolution of inform
 ation technology.  Its impact to statistical research is enormous. This tal
 k attempts to address recent developments and some potential research issue
 s in Business\, Industry and Government (BIG) Statistics\, with special foc
 us on computer experiment and information systems. An overall introduction 
 and review will be given at the beginning\, followed by specific research p
 otentials. For each subject\, the problem will be introduced\, some initial
  results will be presented\, and future research problems will be suggested
 . If time permits\, I will also discuss some recent advances in Search Engi
 ne and RFID study. \n\n Brief Bio Dr. Dennis Lin is a University Distinguis
 hed Professor of Supply Chain Management and Statistics at Penn State Unive
 rsity. His research interests are quality assurance\, industrial statistics
 \, data mining and forecasting. He has published over 100 papers in a wide 
 variety of journals. He currently serves as associate editor for Technometr
 ics\, Statistica Sinica\, Journal of Statistical Theory and Practice\, Jour
 nal of Data Science\, Quality Technology & Quality Management\, Journal of 
 Quality Technology\; and Taiwan Outlook. He is editor of Applied Stochastic
  Models for Business and Industry from 2008. Dr. Lin is an elected fellow o
 f the American Statistical Association (ASA)\, an elected member of Interna
 tional Statistical Institute (ISI)\, an elected fellow of American Society 
 of Quality (ASQ)\, a lifetime member of International Chinese Statistical A
 ssociation (ICSA)\, a fellow of the Royal Statistical Society (RSS)\, and h
 as received the Most Outstanding Presentation Award from SPES\, ASA. He is 
 an honorary chair professor for various universities\, including National C
 hengchi University (Taiwan)\, Remin University of China\, and Fudan Univers
 ity.  He is also the recipient of the 2004 Faculty Scholar Medal Award at P
 enn State University.
SUMMARY:BIG Statistics
DTSTART:20080219T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:41
SEQUENCE:0
DTEND:20080212T160000
UID:2009-11-17T20:55:58-08:00_338841145@limen.stat.ucla.edu
DESCRIPTION:A considerable amount of effort has been recently invested in d
 eveloping a comprehensive  theory for adaptive MCMC.  In comparison\, there
  are fewer adaptive algorithms designed for practical situations. \n\n I wi
 ll discuss review theoretical approaches used for proving convergence of no
 n-Markovian adaptation schemes and will discuss   scenarios for which the o
 riginal adaptive Random-Walk Metropolis is unsuitable.  Alternative adaptiv
 e schemes involving inter-chain and regional adaptation are discussed.  Som
 e of the proposed solutions involve theoretical questions that are still op
 en.
SUMMARY:Learn from Thy Neighbour: Parallel-Chain Adaptive MCMC
DTSTART:20080212T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:42
SEQUENCE:0
DTEND:20080205T160000
UID:2009-11-17T20:55:58-08:00_243594401@limen.stat.ucla.edu
DESCRIPTION:A number of techniques have been developed for dealing with hig
 h dimensional data sets that fall on or near a smooth low dimensional nonli
 near manifold. Such data sets arise whenever the number of modes of variabi
 lity of the data is much smaller than the dimension of the input space\, as
  is often the case for image sequences. Methods that explicitly exploit the
  manifold structure of these high dimensional data sets hold the potential 
 to significantly advance the state of the art in a number of problems in co
 mputer vision and related fields\, although to date few practical algorithm
 s have resulted from the study of non-linear manifolds. Our work seeks to c
 hange that. \n\n In the first part of this talk we present a theoretical ov
 erview of our proposed manifold learning technique. Usually\, manifold lear
 ning is formulated in terms of finding an embedding or 'unrolling' of a man
 ifold into a lower dimensional space. Instead\, we treat it as the problem 
 of learning a representation of a nonlinear\, possibly non-isometric manifo
 ld that allows for the manipulation of novel points. Our resulting algorith
 m\, Locally Smooth Manifold Learning (LSML)\, learns a representation of a 
 manifold or family of related manifolds and can be used for computing geode
 sic distances\, finding the projection of a point onto a manifold\, recover
 ing a manifold from points corrupted by noise\, generating novel points on 
 a manifold\, and more. A crucial characteristic of LSML is the ability to r
 ecover the structure of a manifold in sparsely populated regions and beyond
  the support of the provided data. \n\n In the second part of this talk we 
 show how LSML can be used to learn\, in an unsupervised manner\, the transf
 ormations that a given class of images can undergo\, in order to predict ho
 w a novel image from that class can change. Knowledge of how images of obje
 cts and scenes can change or deform is used in a wide variety of applicatio
 ns in computer vision. Certain classes of transformations\, such as transla
 tion and rotation\, are universally applicable\, but others\, such as the t
 hickening of a handwritten stroke\, are limited to images of a particular c
 lass of objects and can be challenging to model. Through different experime
 nts drawn from images\, videos and tracked features\, we show how to learn 
 transformations that are useful for novel instance generation\, classificat
 ion from few examples\, and feature trajectory de-noising.
SUMMARY:Locally Smooth Manifold Learning
DTSTART:20080205T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:43
SEQUENCE:0
DTEND:20080129T160000
UID:2009-11-17T20:55:58-08:00_456444214@limen.stat.ucla.edu
DESCRIPTION:For the upcoming talk I would like to go over the work we are e
 xploring at The New York Times Research & Development Lab. Here are some of
  the projects I would like to talk about: \n\n - Researching the 'millennia
 l generation' and how they interact with media. \n\n - Newspaper 2.0. We ar
 e researching eInk devices\, flexible displays\, projector mounted mobile d
 evices as possible paper replacements. Some of these devices could affect t
 he way we consume content in the coming years. We are looking at what will 
 make the most sense both for consumers and for delivery in a 24 hour newsro
 om. Will paper as we know it still be around in 10 years or will there be a
  form of digital paper that has replaced the current market. \n\n - GPS. We
  have started to outfit reporters with GPS devices (on an experimental and 
 limited basis so far) that automatically upload their location and mapping 
 info to servers allowing the web team to create maps associated with travel
  and foreign stories the reporters are covering. We are looking to do this 
 with the photographers at The Times also and have mock ups showing how this
  could work in a real world environment with all of our multimedia\, video 
 and reporting content. \n\n - Mobile interaction. We are focusing a lot of 
 our current efforts on mobile devices such as mobile phones\, UMPCs and oth
 er gadgets currently on the market and coming to market in the next 12-24 m
 onths.  Within the mobile space we have already started integrating the new
 spaper with SMS codes that allow the reader to send a text message into The
  Times to get more information pertaining to sections of the paper and spec
 ific articles. In the same respect we have been experimenting with image re
 cognition via the cell phone using 2d barcodes such as semacodes\, QR codes
  and actual images used as visual codes. \n\n - Google Earth. One of the me
 mbers of our group is currently completing a project with Google Earth that
  manages to encode\, tag and geotag our content from our archives and place
  it on a Google Earth layer. This is still in development but I will be abl
 e to show some examples of the work we have done here. \n\n - Shifd.com - S
 hifd was a project that arose from a co-worker\, Michael Young\, and I winn
 ing Yahoo Hack Day in London last year. Shifd allows a user to take content
  between a multitude of devices seamlessly and is currently in production w
 ith a small team we are overseeing. Shifd is a principal we are applying to
  other projects with the R&D Lab that I can touch on. \n\n - The Digital li
 ving room. We are currently about to begin work on a 'personalized' digital
  living room news experience. We have built out a digital living within the
  R&D space that will allow people to walk into the living room and using bl
 uetooth or RFID we will be able to detect who they are\, and their news pre
 ferences\, delivering a personalized newscast to their TV. When the user wa
 lks out of the room -we will be able to detect that they are no longer watc
 hing and send the remaining content of to their mobile device to finish wat
 ching the newscast\, or their car to listen to an audio version. \n\n - And
  also a brief discussion of digital avatars in the media.
SUMMARY:Introduction to Research at the NY Times
DTSTART:20080129T150000
DTSTAMP:20091117T205558
LOCATION:1425 Physics & Astronomy Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:265
SEQUENCE:0
DTEND:20001030T160000
UID:2009-11-17T20:55:58-08:00_227349989@limen.stat.ucla.edu
DESCRIPTION:In recent years statistical extreme value theory has matured to
  such an extent to contribute usefully to the study of substantial real pro
 blems\, particularly in the area of environmental extremes. Examples includ
 e the design of off-shore structures (Coles and Tawn\, 1994) and the study 
 of reservoir flood safety (Anderson and Nadarajah\, 1993). \n\n A fairly co
 mmonly occurring characteristic is that the variables whose extremes are of
  interest are ordered. In hydro-meteorology one thing that is of interest i
 s the dependence of extreme values of <span class='math'>d</span>-hour rain
 fall over a range of values of <span class='math'>d</span>. One approach is
  to fit a multivariate extreme value distribution over that range. If <span
  class='math'>X(d)</span> denotes rainfall aggregated over <span class='mat
 h'>d</span> hours\, and if <span class='math'>d^' >d</span>  then<br/> <spa
 n class='math'>X(d) \leq X(d^') \leq (d^'/d) X(d)</span> for all <span clas
 s='math'>(X(d)\, X(d^'))</span>\, so an order restriction in the multivaria
 te extreme value model is needed. Similar order restrictions arise in the s
 tudy of the joint distributions of large hourly mean wind speeds and large 
 wind gusts. \n\n The aim of this talk is to develop multivariate extremal m
 odels and associated statistical procedures for vector observations whose c
 omponents are subject to an order relationship. We consider only the bivari
 ate case. The results are applied to the joint analysis of rainfall extreme
 s corresponding to different durations.
SUMMARY:Ordered Multivariate Extremes
DTSTART:20001030T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:1
SEQUENCE:0
DTEND:20090602T160000
UID:2009-11-17T20:55:58-08:00_235634438@limen.stat.ucla.edu
DESCRIPTION:We introduce a multiscale approach to crime modeling\, starting
  from stochastic models of criminal-environment interactions. We first disc
 uss geographic profiling\, the estimation of the probability density of whe
 re an offender lives given the locations of a set of observed crimes commit
 ted by the offender. We show how macroscopic\, city-scale density estimates
  that take into account geographic inhomogeneities (housing density\, parks
 \, bodies of water\, etc.) can be derived from the small scale models of cr
 iminal behavior using Bayes' Theorem. We illustrate the implementation of t
 he methodology with "distance to crime" data provided by the LAPD\, where t
 he conditional density is efficiently computed as the solution to a second 
 order elliptic PDE. We then discuss the related problem of crime forecastin
 g\, where crime dynamics on longer time scales must be incorporated into th
 e models. Here self-exciting effects\, similar to those in seismology\, pla
 y an important role and we show how these can be incorporated phenomenologi
 cally\, through ETAS point process models\, or from the small scale models 
 of criminal behavior using nonlinear SPDEs. Using space-time crime data in 
 Los Angeles\, we show how this approach leads to forecasting strategies tha
 t significantly outperform widely used Crime Hotspot Maps.
SUMMARY:Multiscale Methods for Two Classical Problems Arising in Crime Scie
 nce
DTSTART:20090602T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:51
SEQUENCE:0
DTEND:20071106T160000
UID:2009-11-17T20:55:58-08:00_858019795@limen.stat.ucla.edu
DESCRIPTION:The EM algorithm is a special case of a more general algorithm 
 called the MM algorithm first brought to the attention of statisticians by 
 de Leeuw and Heiser in 1977.  Specific MM algorithms often have nothing to 
 do with missing data.  The first M step of an MM algorithm creates a surrog
 ate function that is optimized in the second M step. In minimization\, MM s
 tands for majorize-minimize\; in maximization\, it stands for minorize-maxi
 mize. \n\n This two-step process always drives the objective function in th
 e right direction. Construction of MM algorithms relies on recognizing and 
 manipulating inequalities rather than calculating conditional expectations.
   This survey walks listeners through the construction of several specific 
 MM algorithms. The potential of the MM algorithm in solving high-dimensiona
 l optimization and estimation problems is its most attractive feature.
SUMMARY:An Overview of the MM Algorithm
DTSTART:20071106T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:44
SEQUENCE:0
DTEND:20080122T160000
UID:2009-11-17T20:55:58-08:00_95205703@limen.stat.ucla.edu
DESCRIPTION:This talk is about discovering and modeling previously unspecif
 ied\, recurring themes in a given set of arbitrary images. Given a set of i
 mages containing frequent occurrences of objects from multiple categories\,
  the goal is to learn a compact model of the categories as well as their re
 lationships\, for the purposes of later recognizing/segmenting any occurren
 ces in new images.  Categories are not defined by the user. Also\, whether 
 and where instances of any categories appear in a specific image is not kno
 wn. This problem is challenging\, since it involves the following unanswere
 d questions. What is an object category? What image properties should be us
 ed and how to combine them to discover category occurrences? What is an eff
 icient multicategory representation? \n\n We will examine a methodology\, d
 eveloped during my postdoctoral work in the Beckman Institute at the Univer
 sity of Illinois Urbana-Champaign\, which addresses these questions when ob
 jects are characterized in 2D. A category is defined as a set of 2D objects
  (i.e.\, subimages) sharing photometric\, geometric and topological propert
 ies of their constituent regions (e.g.\, color\, area\, shape\, spatial lay
 out\, and recursive embedding of regions). Each image is represented by a s
 egmentation tree whose nodes correspond to image regions at all natural sca
 les present\, and edges between tree nodes capture the embedding of small r
 egions within larger ones. The nodes contain the associated region properti
 es. The presence of any categories in the image set is then reflected in th
 e frequent occurrence of similar subtrees (i.e.\, 2D objects) within the im
 age segmentation trees. Our methodology is designed to: (1) match image tre
 es to find similar subtrees\; (2) discover categories by clustering similar
  subtrees\, and use the properties of each cluster to learn the model of th
 e associated category\; and (3) learn the grammar of the discovered categor
 ies that compactly captures their recursive definitions in terms of other s
 impler (sub)categories and their relationships (e.g.\, containment\, co-occ
 urrence\, and sharing of simple categories by more complex ones). The model
  structure and region properties associated with model nodes are characteri
 zed by probability density functions (pdf's). Therefore\, learning involves
  considering possible model structures\, and iterative estimation of the pd
 f's for each structure. The resulting minimum-description-length model is u
 sed to simultaneously detect\, recognize and segment all occurrences of the
  learned categories in a new image. This is done by matching the segmentati
 on tree of a new image against the learned category grammar. This matching 
 also provides a semantic explanation of object recognition in terms of the 
 identified subcategories (i.e.\, object parts) along with their spatial rel
 ationships. \n\n The aforementioned methodology can also be used for identi
 fying recurring image themes of more general kind. An example is that of ex
 tracting the stochastically repeating\, elementary parts of image texture\,
  commonly called as texture elements or texels (e.g.\, waterlilies on the w
 ater surface\, fallen leaves on the ground). \n\n BIO: Sinisa Todorovic rec
 eived the joint B.S./M.S. degree with honors in electrical engineering from
  the University of Belgrade\, Serbia\, in 1994. From 1994 until 2001\, he w
 orked as a software engineer in the communications industry. He received hi
 s M.S. and Ph.D. degrees in electrical and computer engineering at the Univ
 ersity of Florida\, Gainesville\, in 2002\, and 2005\, respectively. Since 
 2005\, he holds the position of Postdoctoral Research Associate in the Beck
 man Institute at the University of Illinois Urbana-Champaign\, where he col
 laborates with Prof. Narendra Ahuja. Sinisa's main research interests conce
 rn computer vision and machine learning\, with current focus on unsupervise
 d extraction and representation of spatial structures recurring in images a
 nd video. He is the recipient of Jack Neubauer Best Paper Award 2004 for a 
 publication in IEEE Trans. Vehicular Technology\, and Outstanding Reviewer 
 Award at the International Conf. on Computer Vision (ICCV) 2007. He serves 
 as Associate Editor of Advances in Multimedia.
SUMMARY:What do Those Images Have in Common?
DTSTART:20080122T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:274
SEQUENCE:0
DTEND:19991028T160000
UID:2009-11-17T20:55:58-08:00_944767130@limen.stat.ucla.edu
DESCRIPTION:The Bureau of the Census plans to use statistical sampling as p
 art of its efforts to make Census 2000 as complete and accurate as possible
 . This is done through the Accuracy and Coverage Evaluation (A.C.E.). The A
 .C.E. relies upon a large survey sample and the dual system estimator to me
 asure and correct for coverage errors in the basic enumeration. This talk w
 ill cover the technical design of the A.C.E. It will also discuss some of t
 he legal and policy issues surrounding the use of sampling in Census 2000\,
  and how these issues shaped the current design.
SUMMARY:Sampling and Census 2000
DTSTART:19991028T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:264
SEQUENCE:0
DTEND:20001107T160000
UID:2009-11-17T20:55:58-08:00_135442170@limen.stat.ucla.edu
DESCRIPTION:In this talk\, I will present a common framework for active con
 tours and Mumford-Shah segmentation\, based on the level set method of S. O
 sher and J. Sethian. First I will introduce an active contour model "withou
 t" edges\, based on segmentation and level sets. By this model\, we can det
 ect objects whose boundaries are not necessarily defined by gradient\, as w
 ell as interior contours automatically. Then I will show how this level set
  model can be generalized\, in order to minimize the Mumford-Shah energy fo
 r segmentation\, for piecewise-constant and piecewise-smooth approximations
 . We represent the set of edges via one or more level set functions\, and w
 e propose a new multiphase level set representation\, which has some advant
 ages: we use only <span class='math'>n</span> level set functions to repres
 ent <span class='math'>2^n</span> phases\, and in addition\, we do not have
  the problems of vacuum and overlap\, naturally arising in multiphase probl
 ems. Also\, we will see that triple junctions can be detected and represent
 ed. Finally\, I will show numerical results on various images\, in order to
  validate the algorithm.
SUMMARY:A Common Level Set Framework for Active Contours and Mumford-Shah S
 egmentation
DTSTART:20001107T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:240
SEQUENCE:0
DTEND:20010605T160000
UID:2009-11-17T20:55:58-08:00_162408233@limen.stat.ucla.edu
DESCRIPTION:A system of experimental design is outlined that attempts to en
 compass many of the major work in factorial experimental design of the 20th
  century. The system has four broad branches: (i) regular orthogonal design
 s\, (ii) nonregular orthogonal designs\, (iii) response surface designs\, (
 iv) optimal designs. Regular orthogonal designs include the <span class='ma
 th'>2^{n-k}</span> and <span class='math'>3^{n-k}</span> designs. Major iss
 ues are optimal assignment of factors and interactions via the minimum aber
 ration and related criteria. The problem becomes harder if the factors cann
 ot be treated symmetrically (e.g.\, blocking or split-plot structure\, and 
 robust parameter designs.) Nonregular orthogonal designs were traditionally
  used for factor screening and main effect estimation. They have been shown
  to possess some hidden projection property that allows interactions among 
 a smaller number of factors to be estimated. Response surface designs are u
 sed primarily for exploring parametric surfaces\, while optimal designs are
  chosen to optimize a given criterion based on a specified model. \n\n Rece
 nt work shows that many nonregular designs can be used to screen a large nu
 mber of factors as well as efficiently estimate a quadratic response surfac
 e on projected designs. This shows that the boundary between (ii) and (iii)
  is getting blurred.
SUMMARY:A System of Experimental Design
DTSTART:20010605T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:415
SEQUENCE:0
DTEND:20090929T160000
UID:2009-11-17T20:55:58-08:00_392560661@limen.stat.ucla.edu
DESCRIPTION:Many applications\, ranging from spam filtering to intrusion de
 tection\, are faced with active adversaries. In all these applications\, th
 e future datasets and the training dataset are not from the same population
 \, due to the transformations employed by the adversaries. Hence a main ass
 umption for the existing classification techniques no longer holds and init
 ially successful classifiers will degrade easily. This becomes a game betwe
 en the adversary and the data miner: The adversary modifies its strategy to
  avoid being detected by the current classifier\; the data miner then updat
 es its classifier based on the new threats. We investigate the possibility 
 of an equilibrium in this seemingly never ending game\, where neither party
  has an incentive to change.  Modifying the classifier causes too many fals
 e positives with too little increase in true positives\; changes by the adv
 ersary decrease the utility of the false negative items that are not detect
 ed. We develop a game theoretic framework where equilibrium behavior of adv
 ersarial classification applications can be analyzed\, and provide a soluti
 on for finding an equilibrium point. A classifier's equilibrium performance
  indicates its eventual success or failure. The data miner could then selec
 t attributes based on their equilibrium performance\, and construct an effe
 ctive classifier.
SUMMARY:Adversarial Classification
DTSTART:20090929T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:52
SEQUENCE:0
DTEND:20071030T160000
UID:2009-11-17T20:55:58-08:00_363642350@limen.stat.ucla.edu
DESCRIPTION:Belief Propagation (BP) is a powerful algorithm that computes m
 arginals of functions on a graphical model and is guaranteed to provide a c
 orrect answer when applied to graphs without cycles.  Loopy BP (BP applied 
 to graphs with cycles) while not guaranteed to provide a right answer\, has
  been shown to work very well in many applications.  Some of them are chann
 el code decoding\, turbo-equalization\, joint demodulation and decoding\, i
 mage rendering\, computer vision\, machine learning\, sensor networks\, sta
 tistical physics and many others.  While this work focuses on BP for Low-De
 nsity Parity-Check (LDPC) decoding\, the results may be extended to other a
 pplications. \n\n BP LDPC decoding is an iterative message-passing algorith
 m over the factor graph of the code. The traditional message-passing schedu
 le consists of updating all the variable nodes in the graph\, using the sam
 e pre-update information\, followed by updating all the check nodes of the 
 graph\, again\, using the same pre-update information. Recently several stu
 dies show that sequential scheduling\, in which messages are generated usin
 g the latest available information\, significantly improves the convergence
  speed in terms of number of iterations. Sequential scheduling raises the p
 roblem of finding the best sequence of message updates. We present practica
 l scheduling strategies that use the value of the messages in the graph to 
 find the next message to be updated. Simulation results show that these inf
 ormed update sequences require significantly fewer iterations than standard
  sequential schedules.
SUMMARY:Informed Dynamic Scheduling for Belief-Propagation Decoding
DTSTART:20071030T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:53
SEQUENCE:0
DTEND:20071023T160000
UID:2009-11-17T20:55:58-08:00_438661701@limen.stat.ucla.edu
DESCRIPTION:When modeling inhomogeneous spatial point patterns\, it is ofte
 n of interest to fit a parametric model for the first order intensity funct
 ion (FOIF) of the process in terms of some measured covariates. Estimates f
 or the regression coefficients can be obtained by maximizing a Poisson maxi
 mum likelihood criterion (Schoenberg\, 2005). To estimate the variance of t
 he resulting estimator\, we propose a novel thinned block bootstrap procedu
 re\, which assumes that the point process is second-order intensity reweigt
 hed stationary. To apply this procedure\, only the FOIF but not any high-or
 der terms of the process needs to be estimated. We establish the consistenc
 y of the resulting variance estimator\, and demonstrate its efficacy throug
 h simulations and an application to a real data example.
SUMMARY:A Thinned Block Bootstrap Variance Estimation Procedure for Inhomog
 eneous Spatial Point Patterns
DTSTART:20071023T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:60
SEQUENCE:0
DTEND:20070515T160000
UID:2009-11-17T20:55:58-08:00_328329444@limen.stat.ucla.edu
DESCRIPTION:The National Aeronautics and Space Administration's (NASA) Eart
 h Observing System is a constellation of satellites that collect massive am
 ounts of data about the Earth's systems. These data sets are provided to th
 e research community for climate change studies\, hazard mitigation\, weath
 er prediction and other scientific purposes. However\, the large sizes of t
 hese remote sensing data sets pose problems for a large class of users who 
 seek to understand global\, long-term behavior. \n\n This talk has three pa
 rts. First\, we introduce JPL's Atmospheric Infrared Sounder (AIRS) mission
  and data\, which are typical of NASA's Earth science remote sensing data s
 ets. Second\, we review a non-parametric procedure for compressing these la
 rge data sets in a way that reduces their sizes and complexities while appr
 oximately preserving distributional information. We describe how this has b
 een implemented for the AIRS mission under operational constraints imposed 
 by NASA's data production systems. Finally\, we present some graphical\, ex
 ploratory analyses of the results\, and discuss their implications for clim
 ate studies.
SUMMARY:Understanding Large-scale Structure in Remote Sensing Earth Science
  Data
DTSTART:20070515T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:61
SEQUENCE:0
DTEND:20070424T160000
UID:2009-11-17T20:55:58-08:00_429025500@limen.stat.ucla.edu
DESCRIPTION:Image understanding starts from recognizing the wide variety of
  patterns of image patches at different locations and resolutions. It is th
 erefore useful to understand statistical properties and construct statistic
 al models of image patches of natural scenes. Natural image patches can be 
 roughly classified into three regimes: stochastic textures\, object shapes 
 or textons\, and geometric lines and regions. In this talk\, I will explain
  that these three different regimes can be unified by what we call informat
 ion scaling\, i.e.\, the change of statistical properties of image data ove
 r the change of resolution. More important\, the three regimes of patterns 
 can be modeled within a unified framework of what we call information proje
 ction\, i.e.\, iteratively projecting the current model onto a manifold of 
 distributions to obtain an updated model. \n\n The talk is based on joint w
 ork with  Zhangzhang Si\, Chuck Fleming\, and Song-Chun Zhu.
SUMMARY:Information Projection for Image Patch Modeling
DTSTART:20070424T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:62
SEQUENCE:0
DTEND:20070417T160000
UID:2009-11-17T20:55:58-08:00_708025701@limen.stat.ucla.edu
DESCRIPTION:The ETAS (epidemic type aftershock sequence) model is a special
  branching process for describing the clustering features of earthquake occ
 urrence in both space and time. This talk includes the following issues: \n
 \n 1. The conditional intensity of the ETAS model and its formulation based
  on empirical studies. 2. Estimating simultaneously the spatially non-homog
 eneous background (immigrant) rate\, together with the parameters associate
 d with cluster features. 3. Stochastic declustering procedures to separate 
 earthquakes into background (immigrant) events and  triggered (dependent) e
 vents. 4. Stochastic reconstruction method to discover secrets in the data 
 which are not implied by the model.
SUMMARY:Branching Point Process and Stochastic Clustering
DTSTART:20070417T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:63
SEQUENCE:0
DTEND:20070410T160000
UID:2009-11-17T20:55:58-08:00_550735408@limen.stat.ucla.edu
DESCRIPTION:The effectiveness of antiretroviral therapy is limited by the d
 evelopment of drug resistance.  Rapid and highly error prone replication of
  a large virus population generates mutants that resist the selective press
 ure of drug therapy.  For many antiretroviral drugs\, multiple mutations ar
 e required for resistance\, and an understanding of how these mutations acc
 umulate is key in predicting the development of resistance. \n\n We introdu
 ce a new class of graphical models\, conjunctive Bayesian networks (CBNs)\,
  to describe the evolution of drug resistance. We show that the maximum lik
 elihood CBN can be written down in closed form and give methods for dealing
  with noisy data.  We apply these tools to analyze the development of drug 
 resistance for two protease inhibitors.  Specifically\, we combine the stat
 istical models for the two drugs with a branching process on the viral fitn
 ess landscape in order to estimate the risk that a drug resistant strain wi
 ll develop during treatment with either drug. \n\n This is joint work with 
 Niko Beerenwinkel (Harvard) and Bernd Sturmfels (UC Berkeley).
SUMMARY:Predicting the Risk of Drug Resistance in HIV Using Conjunctive Bay
 esian Networks
DTSTART:20070410T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:416
SEQUENCE:0
DTEND:20091013T160000
UID:2009-11-17T20:55:58-08:00_961987664@limen.stat.ucla.edu
DESCRIPTION:In this talk\, we consider the problem of assessing differentia
 l expression of entire gene sets in complex biological experiments.  We fir
 st propose a latent variable model that directly incorporates the underlyin
 g regulatory network. We then exploit the theory of mixed linear models (ML
 M)\, to develop a general inference framework for analysis of subnetworks\,
  which also accounts for changes in the network structure. We briefly discu
 ss computational issues and apply the proposed method to analyze data from 
 yeast experiments.
SUMMARY:Network Enrichment Analysis in Complex Experiments
DTSTART:20091013T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:422
SEQUENCE:0
DTEND:20091124T160000
UID:2009-11-17T20:55:58-08:00_776950976@limen.stat.ucla.edu
DESCRIPTION:Web applications are becoming increasingly important in offerin
 g software to the user. The main advantage above classical client-based sof
 tware is that new or improved applications can easily be made available to 
 a wide audience.  Furthermore web applications are by design server-based\,
  and therefore make more efficient use of resources and are easier to maint
 ain. \n\n In this presentation will be demonstrated how to create an R web 
 application: a (graphical) web interface for an R function or package.  Som
 e examples of R web applications are demonstrated on my homepage: http://ww
 w.stat.ucla.edu/~jeroen. During the presentation we will take an in depth l
 ook at the design of some of these applications. The presentation will incl
 ude a simple demonstration of how to create such an application\, as well a
 s some theory about the underlying ideas and concepts.
SUMMARY:Web Development with R
DTSTART:20091124T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:66
SEQUENCE:0
DTEND:20070306T160000
UID:2009-11-17T20:55:58-08:00_84982576@limen.stat.ucla.edu
DESCRIPTION:An objective of randomized placebo-controlled vaccine efficacy 
 trials is to evaluate vaccine-induced immune responses as surrogate endpoin
 ts for clinical endpoints such as HIV infection. Dean Follmann introduced a
 ugmented vaccine trial designs for evaluating immunological surrogates\, wh
 erein baseline covariates are used to predict the counterfactual vaccine-in
 duced immune responses of placebo recipients had they been vaccinated\, or 
 uninfected placebo recipients are vaccinated at study-closeout. For these d
 esigns with case-cohort sampling of immune responses\, this talk describes 
 a potential outcomes approach for evaluating the "causal effect predictiven
 ess" of an immune response. Estimated likelihood methods are used for estim
 ating the causal effect predictiveness surface\, which has useful interpret
 ation in terms of how well causal vaccine effects on the immune response pr
 edict causal vaccine effects on the clinical endpoint. The methods incorpor
 ate a model for predicting counterfactual immune responses of placebo recip
 ients\, and apply to binary\, time-to-event\, or quantitative clinical endp
 oints. More generally\, this talk describes a potential outcomes approach f
 or evaluating surrogate endpoints that departs from traditional approaches 
 that are based solely on observable statistical associations. \n\n This is 
 a joint seminar with Biostatistics
SUMMARY:A Potential Outcomes Approach to Quantitating the Predictiveness of
  a Surrogate Endpoint\, with Application to Vaccine Trials
DTSTART:20070306T150000
DTSTAMP:20091117T205558
LOCATION:43-105A Center for Health Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:68
SEQUENCE:0
DTEND:20070223T160000
UID:2009-11-17T20:55:58-08:00_899548719@limen.stat.ucla.edu
DESCRIPTION:I will describe some of our recent efforts in the development o
 f Monte Carlo strategies (both MCMC and SMC) for simulating and optimizing 
 molecular structures. I will illustrate these ideas using examples from Hyd
 rophobic-Hydrophilic (HP) protein model (both 2-D and 3-D) optimization\, p
 rotein side-chain entropy (SCE) estimation\, and near-native structure (NNS
 ) simulations.  By applying the new SMC and MCMC schemes\, we were able to 
 achieve the best results for all the 2-D and 3-D HP structural optimization
  examples we can find in the literature. In particular\, the new approach a
 chieved better results for these HP models than a modified PERM algorithm a
 nd the equi-energy Sampler (Kou et al. 2006). For the SCE and NNS problems\
 , we can characterize accurately many important ensemble properties of NNS 
 and compute efficiently the SCE of a given structural backbone for any give
 n energy function. We also found that widely used pairwise potential functi
 ons behaved surprisingly badly for stabilizing near native protein structur
 es\, and adding a term representing the SCE of the protein can help greatly
  in discriminating true native structures from decoys. \n\n Based on the jo
 int work with Jinfeng Zhang\, and Sam Kou.
SUMMARY:Monte Carlo Methods for Studying Protein Structures
DTSTART:20070223T150000
DTSTAMP:20091117T205558
LOCATION:5200 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:70
SEQUENCE:0
DTEND:20070206T160000
UID:2009-11-17T20:55:58-08:00_771625842@limen.stat.ucla.edu
DESCRIPTION:Humans are quite adept at recognizing objects of the same type 
 despite potentially huge variations in texture\, viewpoint\, and configurat
 ion between these objects.  Current methods of modeling these object classe
 s are woefully lacking\, however\, as they tend to emulate our ability usin
 g rigid models that can model a subset of objects at best\, generally under
  very similar conditions.  In this talk I will present a model for object c
 ategories that achieves greater flexibility than current methods by combini
 ng the hierarchical nature of a stochastic context free grammar with the lo
 cal constraints of a Markov random field.  This allows for a representation
  of objects as hierarchies of parts constrained at each level to enforce co
 nsistency.  Samples drawn from this model show that a higher level "concept
 " of the object is being captured than many methods in use today and I will
  demonstrate using this model for recognition as well as modeling
SUMMARY:A Stochastic Grammar for Object Category Modeling
DTSTART:20070206T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:67
SEQUENCE:0
DTEND:20070227T160000
UID:2009-11-17T20:55:58-08:00_82336226@limen.stat.ucla.edu
DESCRIPTION:We will discuss the computational inference of directed acyclic
  graphical structures given data from experimental interventions. Order-spa
 ce MCMC\, equi-energy sampling\, importance weighting are combined to creat
 e an algorithm for learning Bayesian network structures. \n\n This is based
  on joint work with Byron Ellis.
SUMMARY:Monte Carlo Inference of Bayesian Network Structures
DTSTART:20070227T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:48
SEQUENCE:0
DTEND:20071127T160000
UID:2009-11-17T20:55:58-08:00_608421937@limen.stat.ucla.edu
DESCRIPTION:In this talk (and I use the term loosely) I will present a hand
 ful of (essentially) artworks that could be described (if you were feeling 
 generous) as non-traditional data visualizations. They are non-traditional 
 in the sense that they do not make use of any of the methods or materials w
 e usually associate with our practice (they are not screen-based\, they are
  not printable in any meaningful sense\, R does not figure into their regul
 ar operation in any way ... ). And yet they take their energy from data\; a
 nd become almost portraits (of people\, processes\, institutions) crafted f
 rom data.  I will mainly focus on an artwork completed last week in collabo
 ration with Ben Rubin (EAR Studio) and installed (permanently) in the lobby
  of the new New York Times Building in Manhattan (8th Avenue between 40th a
 nd 41st Streets). Next\, I\'ll try to put the piece in context by discussin
 g some student work from the Design|Media Art class I teach (usually every 
 other year) and some upcoming work that Ben Rubin and I have planned. I wil
 l end with a view (again biased) of what these works mean in terms of the p
 rofessional practice of data analysis and visualization.
SUMMARY:What I did last Summer\, Or from the Halls of the New York Times to
  Graphics by the Hoi Polloi — A Brief Tour of  Recent Data Visualization Wo
 rk
DTSTART:20071127T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:50
SEQUENCE:0
DTEND:20071113T160000
UID:2009-11-17T20:55:58-08:00_14744665@limen.stat.ucla.edu
DESCRIPTION:Missing data are a recurring problem that can cause bias or lea
 d to inefficient analyses.  The development of statistical methods to addre
 ss missingness has been actively pursued in recent years\, including imputa
 tion\, likelihood and weighting approaches (Ibrahim et al.\, JASA\, 2005\; 
 Horton and Kleinman\, TAS\, 2007).  Each approach is considerably more comp
 licated when there are many patterns of missing values and both categorical
  and continuous random variables are involved.  Implementations of routines
  to incorporate observations with incomplete variables in regression models
  are now widely available\, though not commonly used.  We review these meth
 ods in the context of a motivating example from a large health services res
 earch dataset.  Some discussion of the feasibility of sensitivity analyses 
 to the missing at random assumption will also be provided. While there are 
 still limitations to the current implementations\, and additional efforts a
 re required of the analyst\, it is feasible and scientifically desirable to
  incorporate partially observed values as well as undertake sensitivity ana
 lyses to modelling and missingness assumptions.
SUMMARY:Much Ado About Nothing: Methods and Software Implementations to Est
 imate Incomplete Data Regression Models
DTSTART:20071113T150000
DTSTAMP:20091117T205558
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:69
SEQUENCE:0
DTEND:20070213T160000
UID:2009-11-17T20:55:58-08:00_184217550@limen.stat.ucla.edu
DESCRIPTION:Large\, complex data sets are ubiquitous\, the standard now rat
 her than the exception. They present challenging problems of analysis becau
 se of their size and the complexity of their data structures and patterns. 
 One approach is to compute summary statistics at the outset to reduce the c
 omplexity\, but this expedient risks losing important information in the da
 ta.  The goal should be lossless analysis: analyze the data at a level of d
 etail and comprehensiveness that does not sacrifice information. \n\n Achie
 ving lossless analysis of complex data today is immensely challenging.  New
  fundamental approaches and methods are needed for each of the different ar
 eas that come into play in the analysis of the data — databases\, data proc
 essing\, data structures\, statistical models and methods\, machine learnin
 g algorithms\, data visualization\, computational algorithms\, software env
 ironments\, and hardware environments. In fact\, it has never been harder t
 o achieve lossless analysis because complexity has increased faster than ou
 r innovations in these areas. \n\n Nothing serves lossless analysis better 
 than data visualization\, the only practical way to absorb large amounts of
  information in detail. But for today's complex sets we must visualize far 
 larger amounts than in the past. We must be ready to accept large displays 
 each covering tens or even hundreds of screensful (pages).  For a single da
 ta set it is reasonable to have hundreds of such displays.  These displays 
 become a new database produced from the data that is queried and studied. F
 or a display of 500 pages\, we might query and study all or just a few of t
 he pages depending on the task. \n\n Producing\, querying\, and studying a 
 visualization database needs new ideas.  There are different modes of viewi
 ng the many pages and panels per page of a large display\, from slow focuse
 d study to very rapid scans. We need creative interfaces to facilitate the 
 different modes.  We cannot fuss with very large displays\, interacting wit
 h the micro-elements to get them right\, because there is too much\; instea
 d there should be smart automation algorithms that get the large display ri
 ght the first time.  We must consider the physical screen space\, its size 
 and resolution\, to make it work most effectively for the visual study.  We
  need methods of display that result in pre-attentive visual formation of g
 estalts that show instantaneously the relevant patterns in the data.  This 
 necessitates\, strangely\, more displays\, starting with broad brush looks 
 to derivative displays whose redesigns show specific aspects of the broad b
 rush more effectively. It also requires the study of visual perception.
SUMMARY:Visualization Databases for Lossless Analysis of Complex Data Sets
DTSTART:20070213T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:72
SEQUENCE:0
DTEND:20070123T160000
UID:2009-11-17T20:55:58-08:00_613078584@limen.stat.ucla.edu
DESCRIPTION:One of the most important considerations in any hypothesis base
 d fMRI data analysis is to choose the appropriate threshold to construct th
 e activation maps\, which is usually based on p-values. However\, in fMRI d
 ata\, there are three factors which necessitate severe corrections in the p
 rocess of estimating the p-values\; temporal autocorrelation\, multiple com
 parisons problem and the effect of inherent low frequency. A novel and effi
 cient semi-parametric method\, using resampling of normalized spacings of o
 rder statistics\, is introduced to address all the three problems mentioned
  above. The new method makes very few assumptions and demands minimal compu
 tational effort\, unlike other existing resampling methods in fMRI. Results
  using the proposed method are compared with SPM2.
SUMMARY:A Semi-parametric Approach to Estimate the Family-wise Error Rate i
 n fMRI Using Resting-state Data
DTSTART:20070123T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:64
SEQUENCE:0
DTEND:20070403T160000
UID:2009-11-17T20:55:58-08:00_149093809@limen.stat.ucla.edu
DESCRIPTION:Previous work on unsupervised learning has shown that it is pos
 sible to learn feature representations of images\, similar to those employe
 d in the primary visual cortex\, from the statistics of natural images. How
 ever\, such representations are still not readily suited for object recogni
 tion or other high-level visual tasks because they can change drastically a
 s the image changes to due object motion\, variations in viewpoint\, lighti
 ng\, and other factors. In this talk I will describe how bilinear image mod
 els can be used to learn independent representations of the invariances\, a
 nd their transformations\, in natural image sequences. These models provide
  the foundation for learning higher-order feature representations that coul
 d serve as models of higher stages of processing in the cortex\, in additio
 n to having practical merit for computer vision tasks.
SUMMARY:Bilinear Models of Natural Images
DTSTART:20070403T150000
DTSTAMP:20091117T205558
LOCATION:1260 Franz Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:65
SEQUENCE:0
DTEND:20070313T160000
UID:2009-11-17T20:55:58-08:00_977986836@limen.stat.ucla.edu
DESCRIPTION:Over the last twenty eight years\, I have been able to use many
  results from statistics for algorithm development and performance evaluati
 on in my research on image modeling (1979-1985)\, multidimensional signal p
 rocessing (1983-1988)\, image analysis (1981 - ) and computer vision (1985-
  ). I will provide some examples of how useful statistical methods have bee
 n by discussing the following problems: face and gait-based human authentic
 ation in video\, statistical shape theory for human activity recognition an
 d anomaly detection in video\, structure from motion\, shape from shading a
 nd image modeling.
SUMMARY:How I Made a Living in Computer Vision by Using a Few Things I Lear
 nt in Statistics
DTSTART:20070313T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:73
SEQUENCE:0
DTEND:20070116T160000
UID:2009-11-17T20:55:58-08:00_109389014@limen.stat.ucla.edu
DESCRIPTION:The aim of our talk is to present you with an overview of the U
 CLA collections and services available for data and statistical research th
 rough the Young Research Library and ISSR Data Archives.  We will demonstra
 te: \n\n * Data archives website and catalog - http://www.sscnet.ucla.edu/i
 ssr/da/ \n\n * YRL Statistics and Data Resource Guide - http://www.library.
 ucla.edu/yrl/referenc/govinfo/statistics.htm \n\n * ICPSR (Inter-University
  Consortium for Political and Social Research) - http://www.icpsr.umich.edu
 / \n\n * American Factfinder from the US Bureau of the Census - http://fact
 finder.census.gov/ \n\n Speaker contact: \n\n Libbie Stephenson<br/>Directo
 r<br/>UCLA ISSR Data Archives<br/>libbie@ucla.edu \n\n Kris Kasianovitz<br/
 >Young Research Library<br/>Librarian for NGOs\, State\, Local and Canadian
  Government Information<br/>krisk@library.ucla.edu
SUMMARY:Social Science Data Archives and US Census Data
DTSTART:20070116T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:74
SEQUENCE:0
DTEND:20061205T160000
UID:2009-11-17T20:55:58-08:00_147369130@limen.stat.ucla.edu
DESCRIPTION:We consider the first exit time of a nonnegative Harris-recurre
 nt Markov process from the interval [0\,A] as A goes to infinity. We provid
 e a method of proof of asymptotic exponentiality of the first exit time (su
 itably standardized) that does not rely on embedding a regeneration process
 . We also show that under certain conditions the moment generating function
  of a suitably standardized version of the first exit time converges to tha
 t of Exp(1). The results are applied to the evaluation of a distribution of
  run length to false alarm in change-point detection problems. The latter a
 pplication area is discussed in detail.
SUMMARY:Asymptotic Exponentiality of the Distribution of First Exit Times f
 or a Class of Markov Processes with Applications to Change Detection Proble
 ms
DTSTART:20061205T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:75
SEQUENCE:0
DTEND:20061128T160000
UID:2009-11-17T20:55:58-08:00_888560088@limen.stat.ucla.edu
DESCRIPTION:A novel approach to the problem of estimating the heavy-tail ex
 ponent alpha of a distribution is proposed. It is based on the fact that bl
 ock-maxima of size m of  independent and identically distributed data scale
  at a rate of m to the power of 1/alpha. This scaling rate can be captured 
 well by the max-spectrum plot of the data that leads to regression based es
 timators. Consistency and asymptotic normality of these estimators is estab
 lished under mild conditions on the behavior of the tail of the distributio
 n. The results are obtained by establishing bounds on the rate of convergen
 ce of moment-type functionals of heavy-tailed maxima. Practical issues on t
 he automatic selection of tuning parameters for the estimators and correspo
 nding confidence intervals are also addressed. We provide a brief assessmen
 t of the performance of the proposed estimators and illustrate them on sele
 cted real data sets. Some extensions to dependent data would be briefly dis
 cussed.
SUMMARY:Estimating Heavy-tailed Exponents Through Max Self-similarity
DTSTART:20061128T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:80
SEQUENCE:0
DTEND:20061031T160000
UID:2009-11-17T20:55:58-08:00_386837421@limen.stat.ucla.edu
DESCRIPTION:Dependence plays a fundamental role in statistics. In the talk 
 I will introduce a different concept of dependence. The viewpoint provides 
 new insights in the study of complicated random systems. I will also discus
 s relations with nonlinear system theory\, experimental design\, informatio
 n theory\, risk-metrics theory and high dimensional covariance matrices est
 imation.
SUMMARY:Dependence
DTSTART:20061031T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:81
SEQUENCE:0
DTEND:20061017T160000
UID:2009-11-17T20:55:58-08:00_120766581@limen.stat.ucla.edu
DESCRIPTION:Our web page gives techniques to assess the two sample problem 
 (whether or not two different samples are generated by the same probability
  distribution or not). The web page computes normed and rank tests. All tes
 ts are distribution-free because they make no underlying distributional ass
 umptions. The null hypothesis is that the two samples are generated by the 
 same probability distribution. Investigators often make a trade-off between
  robustness and statistical power. Inasmuch as most investigations are abou
 t gathering evidence as opposed to making terminal decisions\, it may make 
 sense to run a battery of tests with different abilities as a sensitivity c
 heck. We find our normed tests to be more powerful in detecting variance ef
 fects\, while performing similarly for mean effects for pseudo-randomly gen
 erated normal and uniform distributions. \n\n Bio: Stephen Kane is an Assis
 tant Professor of finance at Suffolk University in Boston.  He earned his P
 h.D. in mathematics at the Ohio State University. He teaches econometrics\,
  financial forecasting\, money and capital markets\, and financial institut
 ion management.  He is a former Vice President of Public Policy Research at
  Chase Manhattan Bank and former policy analyst at the Office of Management
  and Budget in Washington D.C.  He has consulted for the Federal Reserve Ba
 nk of Chicago and the Fannie Mae Foundation.
SUMMARY:www.normeddistributionfree.org
DTSTART:20061017T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:82
SEQUENCE:0
DTEND:20061010T160000
UID:2009-11-17T20:55:58-08:00_943287108@limen.stat.ucla.edu
DESCRIPTION:Fractional factorial designs are among the most widely used exp
 erimental plans in practice. The main problem in fractional factorial desig
 ns is how to choose and construct efficient or optimal designs. This talk w
 ill review recent advances in the construction of efficient fractional fact
 orial designs\, including regular designs\,  nonregular designs\, blocked d
 esigns\, and supersaturated designs.
SUMMARY:Recent Advances in the Construction of Efficient Large Fractional F
 actorial Designs
DTSTART:20061010T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:78
SEQUENCE:0
DTEND:20061114T160000
UID:2009-11-17T20:55:58-08:00_142438585@limen.stat.ucla.edu
DESCRIPTION:Methods from nonparametric Bayesian statistics are often used t
 o address model selection problems such as determining the number of compon
 ents in a mixture model. Many of these methods are based on the Chinese res
 taurant process (CRP)\, which defines a distribution on assignments of obse
 rvations to components in a way that does not limit the number of component
 s. In this work\, we expand the class of statistical models to which nonpar
 ametric Bayesian methods can be applied by defining a distribution on binar
 y matrices that leaves the number of columns unbounded. This distribution c
 an be used as a prior over matrices of latent features or over bipartite gr
 aphs\, allowing the effective dimensionality of these structures to be dete
 rmined by the data. This distribution has several desirable properties\, an
 alogous to those of the CRP. In particular\, it can be described as the out
 come of a simple sequential process\, which we call the Indian buffet proce
 ss\, and the rows of the resulting matrix are exchangeable. I will outline 
 the basic ideas behind this approach\, as well as some recent applications 
 and extensions. This is joint work with Zoubin Ghahramani and Frank Wood.
SUMMARY:The Indian Buffet Process
DTSTART:20061114T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:83
SEQUENCE:0
DTEND:20061006T150000
UID:2009-11-17T20:55:58-08:00_423340740@limen.stat.ucla.edu
DESCRIPTION:Many statistical users nowadays pay essentially no attention to
  what any of the probabilities they are manipulating mean. There is even an
  'algorithmic school'.  In my experience\, people get quite annoyed when yo
 u ask what they think they are doing. I will try to fan the flames of this 
 into a debate.
SUMMARY:Can You Do Statistics Without Probability?
DTSTART:20061006T130000
DTSTAMP:20091117T205558
LOCATION:C-301 Anderson
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:420
SEQUENCE:0
DTEND:20091006T160000
UID:2009-11-17T20:55:58-08:00_133692357@limen.stat.ucla.edu
DESCRIPTION:Remote sensing data are often sparse relative to the space-time
  domains of geophysical processes\, so data from different instruments are 
 used in conjunction with one another to take advantage of their complementa
 ry coverage. Yet\, there is no comprehensive data fusion methodology for co
 mbining them in a principled\, rigorous way. Remote sensing data are often 
 massive\, with incompatible support\, and subject to biases. We approach th
 is problem from a statistical point of view\, and aim to estimate the condi
 tional means of true but not directly observed processes from noisy and pos
 sibly biased data  realizations\, and also to estimate the uncertainties of
  these estimates. This dissertation proposes an optimal fusion methodology 
 that scales linearly with data size\, and resolves change of support and bi
 ases through a spatial statistical framework. \n\n This methodology is base
 d on Fixed-ranked Kriging (FRK)\, a variant of kriging that uses a special 
 class of covariance functions for spatial interpolation of a single\, massi
 ve input dataset. This simplifies the computations needed to calculate the 
 kriging means and prediction errors. We extend the FRK framework to the cas
 e of two or more massive input datasets.  The methodology does not require 
 assumptions of stationary or isotropy\, making it appropriate for a wide ra
 nge of geophysical processes. The method also accounts for change of suppor
 t\, allowing estimation of the point-level covariance functions from aggreg
 ated data\, and prediction to point-level locations. This new methodology i
 s applied to aerosol optical depth data from two remote sensing instruments
 \, where the total data size is on the order of tens of thousands.
SUMMARY:Spatial Statistical Data Fusion for Remote Sensing Applications
DTSTART:20091006T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:421
SEQUENCE:0
DTEND:20091027T160000
UID:2009-11-17T20:55:58-08:00_435391888@limen.stat.ucla.edu
DESCRIPTION:"The Human Face Project" is a short film documenting an effort 
 at Walt Disney Feature Animation to track and animate human facial performa
 nce\, which was shown in the SIGGRAPH 2002 Electronic Theater. This present
 ation will outline the techniques developed in this project\, and demonstra
 ted in that film. The face tracking system we developed is an example of mo
 del-based computer vision\, and exploits the detailed degrees of freedom of
  a geometric face model to confine the space of solutions. Optical flow and
 /or successive rerendering of the model are employed in an optimization loo
 p to converge on model parameter estimates. The structure of the model perm
 its very principled mapping of estimated expressions to different targets. 
  Of critical importance for media applications is the handling of details b
 eyond the resolution or degrees of freedom of the tracking model. We descri
 be behavioral modeling expedients for realizing these details in a plausibl
 e way. \n\n Biosketch: \n\n Lance J. Williams is an Academy Award and Steve
 n A. Coons Award winning graphics researcher who made major contributions t
 o texture map prefiltering\, shadowing algorithms\, facial animation\, and 
 image-based rendering. \n\n Prior to his current position as principal rese
 archer at Nokia Research Center Hollywood\, Lance was a software engineer f
 or Google Earth\, senior scientist at Applied Minds\, Chief Scientist at Wa
 lt Disney Feature Animation\, senior software engineer at DreamWorks SKG\, 
 and member of technical staff in Apple's Advanced Technology Group\, where 
 he contributed to QuickTime VR. \n\n He graduated from the University of Ka
 nsas in 1972\, and attended graduate school at the University of Utah.  He 
 was awarded a Ph.D. in computer science from the University of Utah in 2000
 .
SUMMARY:Disney Human Face Project &mdash\; Capture and Transfer of Facial D
 ynamics 
DTSTART:20091027T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:85
SEQUENCE:0
DTEND:20060530T160000
UID:2009-11-17T20:55:58-08:00_302215465@limen.stat.ucla.edu
DESCRIPTION:In atmospheric science the understanding of the evolution and i
 nteraction of turbulence systems has been the focus of many studies. In thi
 s talk we describe methods for (i) identifying and (ii) tracking such turbu
 lence structures captured in image sequences.  For the identification probl
 em\, we propose a multi-resolution statistical model that utilizes the non-
 decimated discrete wavelet transform\, while for the tracking problem\, two
  approaches are investigated.  In the first approach a bivariate AR time se
 ries is used to model the trajectory of each vortex\, and in the second app
 roach the trajectory is modeled by integrated Brownian motion.  A main appl
 ication of this work is the automatic tracking of storm activities observed
  in remote sensing images.
SUMMARY:Identifying and Tracking Turbulence Structures
DTSTART:20060530T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:89
SEQUENCE:0
DTEND:20060517T153000
UID:2009-11-17T20:55:58-08:00_502798305@limen.stat.ucla.edu
DESCRIPTION:It is becoming increasingly popular to use machine learning met
 hods in computer vision\, especially for tasks involving object detection a
 nd tracking. One of the main hurdles at this time is the collection of trai
 ning data. The amount of manual labor required for accurately labeling a su
 fficient number of training examples quickly becomes prohibitive. The term 
 "active learning" describes a set of methods whereby the learning algorithm
  chooses the examples the human should label. \n\n In this talk I will pres
 ent a system\, called "Seville" for learning to detect pedestrian from a ca
 mera mounted on the front of a car. I will demonstrate the use of active le
 arning and the theoretical underpinning of the algorithm. \n\n This is join
 t work with Yotam Abramson and Rafi Pelossof. \n\n Yoav Freund is a profess
 or of Compuer Science in the University of California/San Diego. His work i
 s in the area of machine learning\, computational statistics and their appl
 ications. Freund and Schapire invented the Adaboost algorithm for which the
 y recieved the Godel Prize in 2003 and the Kanellakis prize in 2004.
SUMMARY:Active Learning For Object Detection
DTSTART:20060517T143000
DTSTAMP:20091117T205558
LOCATION:4760 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:84
SEQUENCE:0
DTEND:20060606T160000
UID:2009-11-17T20:55:58-08:00_488947148@limen.stat.ucla.edu
DESCRIPTION:A Neyman-Scott process is a point process consist of clusters\,
  where the locations of the members in each cluster are i.i.d. random aroun
 d its unobservable cluster center. In the general theory of point processes
 \, the conditional intensity function (the occurrence rate of a point at a 
 particular time given all the previous history) plays a key role in the eva
 luation of the likelihood. This talk gives the expression of the conditiona
 l intensity and the likelihood function given observation in an finite regi
 on. With the introduction of the filtering technique\, we can also evaluate
  the locations of the cluster centers.
SUMMARY:Neyman-Scott Processes: Conditional Intensity\, Likelihood and  Det
 ection of Cluster Centers
DTSTART:20060606T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:56
SEQUENCE:0
DTEND:20070726T160000
UID:2009-11-17T20:55:58-08:00_916418694@limen.stat.ucla.edu
DESCRIPTION:This is joint work with Mark Heiler (University of Konstanz). A
  regression cross covariance and cross spectrum for multivariate determinis
 tic trend functions is defined. This is a nonparametric multivariate extens
 ion of the regression spectrum used by Grenander and Rosenblatt to obtain a
 symptotic results in the context of parametric regression. The usefulness o
 f the nonparametric regression spectrum goes far beyond a purely mathematic
 al device. It can be used as a data analytical tool to identify common freq
 uencies and lead-lag effects in multivariate time series with strong determ
 inistic components. Alternative approaches to estimating the regression spe
 ctrum\, as well as its modulus and phase\, include a) wavelet thresholding\
 , b) Fourier transform and c) sample cross-correlation. Asymptotic properti
 es of a)-c) are derived\, and algorithmic issues are discussed. The methods
  are illustrated by examples from biology\, medicine and environmental scie
 nces.
SUMMARY:On Dependence and Phase Shifts in Multivariate Time Series
DTSTART:20070726T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:77
SEQUENCE:0
DTEND:20061117T160000
UID:2009-11-17T20:55:58-08:00_996538240@limen.stat.ucla.edu
DESCRIPTION:Graphical Markov models have some advantages over other multiva
 riate statitical models. They permit to model stepwise data generating proc
 esses with and without interventions and to work out consequences of a larg
 e model for subsets of variables. Such implications can then be compared wi
 th available background knowledge or may be used to judge seemingly inconsi
 stent results in similar studies.  Two recently developed matrix operators\
 , one for real-valued matrices and one for binary matrices are most useful 
 for understanding and deriving properties and implications of graphical Mar
 kov models. \n\n Wermuth\, N.\, Wiedenbeck\, M. and Cox\, D.R.] (2006). Par
 tial inversion for linear systems and partial closure of independence graph
 .  BIT\, Numerical Mathematics. (Online preprint has appeared in BIT) \n\n 
 Wermuth\, N.\, Cox\, D.R.  and Marchetti\, G.M.] (2006). Covariance chains.
  Bernoulli\, 12\, 841-862. \n\n Wermuth\, N. (2005). Graphical  chain model
 s. In: Encyclopedia of  Behavioral Statistics\, II\, B. Everitt and David C
 . Howell (eds.) Wiley\,  Chichester\, 755-757.
SUMMARY:Partial Inversion and Partial Closure of Paths on Graphs: Two Matri
 x Operators to Study Properties of Large Systems Generated Over Graphs
DTSTART:20061117T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:86
SEQUENCE:0
DTEND:20060524T170000
UID:2009-11-17T20:55:58-08:00_489182041@limen.stat.ucla.edu
DESCRIPTION:Biomarkers are conventionally expected to show a direct associa
 tion with the outcome of interest.  If the patient population is heterogene
 ous\, biomarker/outcome associations are diluted by confounding factors\, s
 o there may be few if any markers showing a uniform predictive relationship
  to the outcome.  I will describe work in progress on uncovering a latent p
 opulation structure best revealing biomarkers that are differentially predi
 ctive in distinct subsets of patients.  The latent population structure is 
 characterized in terms of a univariate continuum of scores assigned to the 
 subjects. Subjects with similar scores have similar outcome/marker associat
 ions\, while subjects with very different scores tend to have quite differe
 nt outcome/marker relationships.  I will demonstrate the approach using a s
 et of lung cancer gene expression measurements from microarrays.
SUMMARY:Identification of Differential Biomarker/Outcome Associations
DTSTART:20060524T160000
DTSTAMP:20091117T205558
LOCATION:5272 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:88
SEQUENCE:0
DTEND:20060518T153000
UID:2009-11-17T20:55:58-08:00_176629012@limen.stat.ucla.edu
DESCRIPTION:I will suggest that experiments-when feasible-offer more reliab
 le evidence on causation than observational studies\; this is not to gainsa
 y the contribution to knowledge from observation. I will also suggest that 
 experiments should be analyzed as experiments not observational studies. Fo
 r instance\, a simple comparison of rates might be just the right tool. So 
 far as time permits\, I will discuss models for experimental data\, the int
 ention-to-treat principle\, and the effect of treatment on the treated. I h
 ave two related papers on the web\, plus a handout: \n\n http://www.stat.be
 rkeley.edu/~census/oxcause.pdf<br/> http://www.stat.berkeley.edu/~census/ne
 yreg.pdf<br/> http://www.stat.berkeley.edu/~census/modelobs.pdf
SUMMARY:Are Experiments Better Than Observational Studies? If So\, How Shou
 ld We Analyze Them?
DTSTART:20060518T143000
DTSTAMP:20091117T205558
LOCATION:1200 IPAM Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:79
SEQUENCE:0
DTEND:20061107T160000
UID:2009-11-17T20:55:58-08:00_541668740@limen.stat.ucla.edu
DESCRIPTION:In many important statistical applications\, the number of vari
 ables or parameters is much larger than the number of observations.  In rad
 iology and biomedical imaging for instance\, one is typically able to colle
 ct far fewer measurements about an image of interest than the unknown numbe
 r of pixels. Examples in functional MRI and tomography immediately come to 
 mind. Other examples of high-dimensional data in genomics\, signal processi
 ng and many other fields abound.  In the context of mulitple linear regress
 ion for instance\, a fundamental question is whether it is possible to esti
 mate a vector of parameters of size p from a vector of observations of size
  n when n is far smaller than p. This  seems a priori hopeless. \n\n This t
 alk introduces a new estimator\, dubbed the Dantzig selector in honor of th
 e late George Dantzig as it invokes linear programming\, and which enjoys r
 emarkable statistical properties. Suppose that the data or design matrix ob
 eys a uniform uncertainty principle and that the true parameter vector is s
 ufficiently sparse or compressible which roughly guarantees that the model 
 is identifiable. Then the estimator achieves an accuracy which nearly equal
 s that one would achieve with an oracle that would supply perfect informati
 on about which coordinates of the unknown parameter vector are nonzero and 
 which were above the noise level. Our results connect with the important mo
 del selection problem. In effect\, the Dantzig Selector automatically selec
 ts the subset of covariates with nearly the best predictive power\, by solv
 ing a convenient linear program. \n\n The results are parts of a larger bod
 y of work perhaps best known as Compressive Sampling or Compressed Sensing.
  If time allows\, I will discuss connections with other fields such as sign
 al processing and coding theory.
SUMMARY:The Dantzig Selector:  Statistical Estimation When p is Larger Than
  n
DTSTART:20061107T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:417
SEQUENCE:0
DTEND:20091110T160000
UID:2009-11-17T20:55:58-08:00_772491092@limen.stat.ucla.edu
DESCRIPTION:Prediction markets are used in real life to predict outcomes of
  interest such as presidential elections. In this work we introduce a mathe
 matical theory for Artificial Prediction Markets for the purpose of supervi
 sed classifier aggregation and probability estimation. In this direction\, 
 we bring the following contributions. First\, we derive the market equation
 s starting from the total budget conservation condition\, show the market p
 rice uniqueness and give efficient algorithms for computing it. Second\, we
  present a method for classifier aggregation using the market price as the 
 estimated conditional probability given the evidence presented to the marke
 t participants though a feature vector x. We also show how to train the mar
 ket in a supervised manner\, updating the participants budgets based on the
  market price and the amount they bet. Third\, we introduce classifier spec
 ialization as a new type of differentiating characteristic of classifiers. 
 Finally\, we present an application to multi-class classification using ran
 dom decision rules as specialized classifiers and show that the prediction 
 market consistently outperforms Random Forest on an array of datasets with 
 Bayes errors ranging from 0 (very easy) to 0.5 (impossible).
SUMMARY:Supervised Aggregation of Classifiers using Artificial Prediction M
 arkets
DTSTART:20091110T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:423
SEQUENCE:0
DTEND:20091117T160000
UID:2009-11-17T20:55:58-08:00_471658752@limen.stat.ucla.edu
DESCRIPTION:I will discuss a notion of visual information as complexity not
  of the raw data\, but of the images after the effects of nuisance factors 
 such as viewpoint and illumination are discounted. It is rooted in ideas of
  J. J. Gibson\, and stands in contrast to traditional information as entrop
 y or coding length of the data regardless of its use\, and regardless of th
 e nuisance factors affecting it. Its computation is made possible by a rece
 nt characterization of the set of images modulo viewpoint and contrast chan
 ges\, that induce group (invertible) transformations on the domain and rang
 e of the image. The non-invertibility of nuisances such as occlusion and qu
 antization induces an "information gap"  that can only be bridged by contro
 lling the data acquisition process. Measuring visual information entails ea
 rly vision operations\,  tailored to the structure of the nuisances so as t
 o be "lossless" with respect to visual decision and control tasks (as oppos
 ed to data transmission and storage tasks implicit in traditional informati
 on theory). I illustrate these ideas on visual exploration\, whereby a "Sha
 nnonian Explorer" navigates unaware of the structure of the physical space 
 surrounding it\, while a "Gibsonian Explorer" is guided by the topology of 
 the environment\, despite measuring only images of it\, without performing 
 3D reconstruction. This operational definition of visual information sugges
 ts desirable properties that a visual representation should possess to best
  accomplish vision-based decision and control tasks.
SUMMARY:Shannon Meets Gibson: Actionable Information and the Link Between V
 ision and Control
DTSTART:20091117T150000
DTSTAMP:20091117T205558
LOCATION:4660 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:90
SEQUENCE:0
DTEND:20060509T160000
UID:2009-11-17T20:55:58-08:00_539452353@limen.stat.ucla.edu
DESCRIPTION:As the rate of enrollment in the undergraduate classes increase
 s\, the College of Letters and Science is looking into blended instruction\
 , combining the regular methods of teaching and technology\, as a mean of e
 nhancing the quality of teaching\, learning\, and assessment. \n\n Last yea
 r we presented a pilot study that related to using blended instruction in t
 eaching introductory statistics to a class that consisted of 32 students. T
 he results indicated that the pilot was successful and we had reached the m
 ajor goals of the restructured Statistics 10 including helping students gen
 erate their own knowledge\, enhancing upper level thinking\, introducing st
 atistics as a science of data\, and creating an interactive environment tha
 t enhanced student-student\, student-TA\, and student-instructor interactio
 n. \n\n In this presentation\, we will discuss the conclusions that have re
 sulted from an experimental study that was designed to examine the effectiv
 eness of blended instruction on teaching introductory statistics to classes
  of 100 students. This will include quantitative results regarding the comp
 arison of the control (old Stat 10) and the experimental group (restructure
 d Stat 10) on acquisition of knowledge\, attitudes examined in the pilot st
 udy\, as well as qualitative results covering student comments\, teaching a
 ssistant comments\, and classroom observations. In addition to positive fin
 ding with respect to the acquisition of statistical knowledge\, the positiv
 e attitudinal results that are in synch with what was reported last year wi
 ll be discussed.
SUMMARY:Examining the Effectiveness of Blended Instruction on Teaching Intr
 oductory Statistics
DTSTART:20060509T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:93
SEQUENCE:0
DTEND:20060418T160000
UID:2009-11-17T20:55:58-08:00_826679458@limen.stat.ucla.edu
DESCRIPTION:We describe an efficient sequential Monte Carlo method for samp
 ling multiway tables with given constraints\, which can be used to approxim
 ate exact conditional inference on contingency tables.  An essential featur
 e of our new method is that it samples table entries sequentially according
  to an appropriate proposal distribution.  The sequential sampling approach
 &mdash\;divides and conquers&mdash\;the difficult task of finding an approp
 riate proposal distribution for a multiway table with complex constraints. 
 Computational commutative algebra is used to provide conditions that guaran
 tee that our method has certain good properties. We apply our method to a r
 ange of examples from social and medical sciences to demonstrate its effici
 ency in real problems.
SUMMARY:Sampling for Conditional Inference on Multiway Tables
DTSTART:20060418T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:95
SEQUENCE:0
DTEND:20060407T160000
UID:2009-11-17T20:55:58-08:00_718557187@limen.stat.ucla.edu
DESCRIPTION:For many years\, computer vision researchers have worked hard c
 hasing the illusive goals\, such as\, can the robot find a boy in the scene
 ? or can your vision system automatically segment the cat from the backgrou
 nd?  These tasks require a lot of prior knowledge and contextual informatio
 n. How to incorporate prior knowledge and contextual information into visio
 n systems\, however\, is very challenging. In this talk\, we propose that m
 any difficult vision tasks can only be solved with interactive vision syste
 ms\, by combining powerful and real-time vision techniques with intuitive a
 nd clever user interfaces.  We will show two interactive vision systems we 
 developed recently\, Lazy Snapping (Siggraph 2004) and Image Completion (Si
 ggraph 2005). Lazy Snapping cuts out an object from a picture using graph c
 ut\, while Image Completion recovers unknown region in a picture with belie
 f propagation. A key element in designing such interactive systems is how w
 e model the user's intention using conditional probability (context) and li
 kelihood associated with user interactions. Given how ill-posed most image 
 understanding problems are\, it is proposed that interactive computer visio
 n is the paradigm we should focus today's vision research on. \n\n A quick 
 overview will also be given of Microsoft Research Asia\, "the world's hotte
 st computer lab" (MIT Technology Review\, June 2004)\, including activities
  in basic research\, technology transfer\, product incubation\, technology 
 licensing\, university relations and internship program. \n\n Biosketch:<br
 /> Harry Shum is a Distinguished Engineer of Microsoft Corporation.  He rec
 eived his Ph.D. in robotics from the School of Computer Science\, Carnegie 
 Mellon University.  After he graduated\, he worked as a researcher at Micro
 soft Research Redmond.  In 1999\, he moved to Microsoft Research Asia (Beij
 ing\, China) where he is currently the Managing Director. \n\n His research
  interests include computer vision\, graphics\, human computer interaction\
 , statistical learning and robotics.  He is on the editorial boards for IEE
 E Transactions on Pattern Analysis and Machine Intelligence (PAMI)\, and In
 ternational Journal of Computer Vision (IJCV).  He is the Program Co-Chair 
 of Eleventh International Conference on Computer Vision (ICCV 2007 Brazil).
   He is a Fellow of IEEE.
SUMMARY:Prior\, Context and Interactive Computer Vision
DTSTART:20060407T150000
DTSTAMP:20091117T205558
LOCATION:1200 IPAM Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:96
SEQUENCE:0
DTEND:20060316T160000
UID:2009-11-17T20:55:58-08:00_441330633@limen.stat.ucla.edu
DESCRIPTION:Cis-regulatory modules composed of multiple transcription facto
 r binding sites control gene expression in eukaryotic genomes. We propose a
  hierarchical mixture approach to model the cis-regulatory module structure
 . Based on the model\, a new de novo motif-module discovery algorithm\, Cis
 Module\, is developed for the Bayesian inference of module locations and wi
 thin-module binding sites. We illustrate the use of CisModule by its applic
 ation to the discovery of a novel tissue-specific regulatory module in Cion
 a savignyi. In addition\, comparative genomic studies show that regulatory 
 elements are more conserved across species due to evolutionary constraints.
  Thus we further extend our approach to combine both module structure and c
 ross-species orthology in motif discovery. We use a hidden Markov model (HM
 M) to capture the module structure in each species and couple these HMMs th
 rough multiple-species alignment. Our new method has been tested on both si
 mulated and biological data sets\, where we observe improvement in motif di
 scovery compared to other widely used computational methods.
SUMMARY:Detecting Cis-Regulatory Modules and Motifs by Modeling Correlated 
 Structures in Genomic Sequences
DTSTART:20060316T150000
DTSTAMP:20091117T205558
LOCATION:5436 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:98
SEQUENCE:0
DTEND:20060309T160000
UID:2009-11-17T20:55:58-08:00_900661493@limen.stat.ucla.edu
DESCRIPTION:A Bayesian approach to statistical analysis is becoming an incr
 easingly valuable modeling tool across many different fields. This talk wil
 l illustrate the use of a Bayesian framework in the context of cognitive re
 search investigating how humans draw strong inferences from noisy and spars
 e data.  I will present computational models that have been developed for t
 wo different domains in which inference is critical\, motion perception and
  causal reasoning. The picture emerging from this work is that a key basis 
 for human inference is the use of generic priors\, i.e. general assumptions
  people make about the way the world works\, which then guide their learnin
 g and inference from observed data. I will sketch broader implications\, in
 cluding future directions for applying Bayesian modeling.
SUMMARY:Bayesian Inference in Cognitive Science
DTSTART:20060309T150000
DTSTAMP:20091117T205558
LOCATION:5436 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:100
SEQUENCE:0
DTEND:20060302T160000
UID:2009-11-17T20:55:58-08:00_957026802@limen.stat.ucla.edu
DESCRIPTION:Clustering methods attempt to find natural groupings in data us
 ing observed features. It is not usually known a priori which observed feat
 ures will provide meaningful information about a clustering. Thus learning 
 which features are important is a fundamental problem for all clustering me
 thods. In this talk\, I will discuss learning specifically for a method cal
 led spectral clustering. Spectral clustering is a technique which uses pair
 wise similarities in order to estimate the cluster membership of the data p
 oints.  I will address the problem of learning these similarities as a func
 tion of observed features.  Learning the importance of each of the observed
  features to the clustering can be viewed as learning parameters correspond
 ing to each feature\, that weigh how heavily it should be used when constru
 cting similarities.  I will formulate an objective for learning these featu
 re weights\, which balances a clustering quality term\, the gap\, and a sta
 bility term\, the eigengap. I will describe an algorithm which finds the fe
 ature weights optimizing the objective function in a supervised setting as 
 well as how this can be extended to the unsupervised case.  I will also pre
 sent the results from experiments which confirm the validity of this approa
 ch.
SUMMARY:Learning in Spectral Clustering
DTSTART:20060302T150000
DTSTAMP:20091117T205558
LOCATION:5436 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:101
SEQUENCE:0
DTEND:20060228T160000
UID:2009-11-17T20:55:58-08:00_350969904@limen.stat.ucla.edu
DESCRIPTION:The literature on adaptive design of clinical trials and sample
  size adjustment is dominated by two-stage tests that use conditional power
  to determine the second-stage sample size.  We present a new three-stage a
 pproach that uses the stopping rules of efficient fully-sequential tests to
  guide its choice of sample size.  These tests control the type I error pro
 bability\, have a fixed maximum sample size\, and easily generalize to high
 er dimensions.  We discuss the asymptotic optimality of these three-stage t
 ests\, the necessity of allowing for three stages\, and the inefficiency of
  the currently used conditional power tests.  Simulations show favorable fi
 nite-sample performance compared to conditional power tests and Stein's two
 -stage procedure. \n\n Biosketch: Jay Bartroff\, PhD\, earned a BA from U.C
 . Berkeley in 1998 and a PhD from California Institute of Technology in 200
 4.  From 1998 until 2000 he was a research associate in the Medical Physics
  and Imaging Department at Cedars-Sinai Medical Center.  He is currently NS
 F Postdoctoral Fellow in the Statistics Department at Stanford University. 
  His research concerns sequential analysis\, clinical trial design\, design
  of experiments\, and imaging.
SUMMARY:Efficient Three-stage Adaptive Designs for Clinical Trials
DTSTART:20060228T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:103
SEQUENCE:0
DTEND:20060214T160000
UID:2009-11-17T20:55:58-08:00_344348150@limen.stat.ucla.edu
DESCRIPTION:Real-parameter Evolutionary Monte Carlo algorithm (henceforth r
 eferred to as EMC) has been proposed as an effective method not only for sa
 mpling from multi-modal distributions but also for stochastic optimization.
  The authors (Liang and Wong\, 2001) have shown by various well-chosen exam
 ples that it works better than parallel tempering. It has a certain ability
  to learn from the past and it improves mixing by sampling along a temperat
 ure ladder. We introduce four new moves which equip the algorithm with more
  learning capability while improving the quality of the samples (as measure
 d by total auto-correlation time) produced from the target distribution. Se
 condly\, we present a new easy to implement strategy for determining the te
 mperature range and construction of the temperature ladder. We illustrate t
 he above techniques through various examples. Thirdly\, we generalize a res
 ult originally proved in (Gilks\, Roberts\, George\, 1994) about the validi
 ty of the conditional (or line) sampling step in the context of the snooker
  algorithm which is a type of move involved in the EMC. \n\n We also consid
 er the problem of clustering a group of observations according to some obje
 ctive function (e.g. K-means clustering\, variable selection) or according 
 to a posterior density (e.g. posterior from a Dirichlet Process prior) of c
 luster indicators. We cast both kinds of problems in the framework of sampl
 ing for cluster indicators. So far\, Gibbs sampling\, "split-merge" Metropo
 lis-Hasting algorithm and various modifications of these have been the basi
 c tools used for sampling in this context. We propose a new population base
 d MCMC approach\, in the same vein as parallel tempering. We introduce thre
 e new "crossover moves" (based on swapping and reshuffling sub-clusters ins
 tersections) which make such an algorithm very efficient with respect to In
 tegrated Autocorrelation Time (IAT) of various relevant statistics and also
  with respect to the ability to escape from local modes. We call this new a
 lgorithm Population Based Clustering (PBC) algorithm. We apply PBC algorith
 m to motif clustering\, Beta mixture of Bernoulli clustering and a Bayesian
  Information Criterion (BIC) based variable selection problem. We also disc
 uss clustering of mixture of Normals and compare the performance PBC algori
 thm as a stochastic optimizer with K-means clustering. \n\n Gopi Goswami is
  lecturer and post-doc in the Department of Statistics\, Harvard University
 .
SUMMARY:On Population Based Markov Chain Monte Carlo Methods
DTSTART:20060214T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:105
SEQUENCE:0
DTEND:20060131T160000
UID:2009-11-17T20:55:58-08:00_859253297@limen.stat.ucla.edu
DESCRIPTION:Many double-blind placebo-controlled randomized experiments wit
 h active drugs suffer from complications beyond simple noncompliance. First
 \, the compliance with assigned dose is often partial\, with patients takin
 g only part of the assigned dose\, whether active or placebo. Second\, the 
 blinding may be imperfect in the sense that there may be detectable positiv
 e or negative side-effects of the active drug\, and consequently\, simple c
 ompliance has to be extended to allow different compliances to active drug 
 and placebo.  Efron and Feldman (1991) presented an analysis of such a situ
 ation and discussed inference for dose-response from the non-randomized dat
 a in the active treatment arm\, which stimulated active discussion\, includ
 ing concerning the role of the intention-to-treat principle in such studies
 .  Here\, we formulate the problem within the principal stratification fram
 ework of Frangakis and Rubin (2002)\, which adheres to the intention-to-tre
 at principle\, and we present a new analysis of the Efron-Feldman data with
 in this framework. Moreover\, we describe precise assumptions under which d
 ose-response can be inferred from such non-randomized data\, which seem deb
 atable in the setting of this example. Although this article only deals in 
 detail with the specific Efron-Feldman data\, the same framework can be app
 lied to various circumstances in both natural science and social science.
SUMMARY:Principal Stratification for Causal Inference with Extended Partial
  Compliance
DTSTART:20060131T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:106
SEQUENCE:0
DTEND:20060124T160000
UID:2009-11-17T20:55:58-08:00_667579433@limen.stat.ucla.edu
DESCRIPTION:There is a wealth of cosmological information encoded in the sp
 atial power spectrum of temperature anisotropies of the cosmic microwave ba
 ckground.  The sky\, when viewed in the microwave\, is very uniform\, with 
 a nearly perfect blackbody spectrum at 2.7 degress.  Very small amplitude b
 rightness fluctuations (to one part in a million!!) were detected by the CO
 BE satellite\, and have now been mapped by ground\, balloon\, and satellite
  instruments to a spatiall resolution smaller than 1 degree.  These brightn
 ess fluctuations trace small density perturbations in the early universe (r
 oughly 300\,000 years after the Big Bang)\, which later grow through gravit
 ational instability to the large scale structure seen in redshift surveys. 
  The details of the physics in the early universe leaves a tell-tale signat
 ure on the statistical structure of hot and cold spots\, with more details 
 of the physics encoded at sub-angular degree spatial scales.  With the push
  to map the microwave sky at higher spatial resolution has come a flood of 
 data\, with maps containomg millions of pixels observed at several differen
 t frequencies (from 30 to 900 GHz)\, all with slightly different resolution
 s and noise properties.  The resulting analysis challenge is to estimate\, 
 and quantify our uncertainty in\, the spatial power spectrum of the cosmic 
 microwave background given the complexities of 'missing data'\, foreground 
 emission\, and complicated instrumental noise.  In this talk\, I will discu
 ss a Bayesian formulation of this problem\, discuss a Gibbs sampling approa
 ch to numerically sampling from the Bayesian posterior\, and the applicatio
 n of this approach to the first-year data from the Wilkinson Microwave Anis
 otropy Probe.  I will also comment on recent algorithmic developments for t
 his approach to be tractabe for the even more massive data set to be retune
 d from the Planck satellite.
SUMMARY:Bayesian Analysis of the Power Spectrum of the Cosmic Microwave Bac
 kground
DTSTART:20060124T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:107
SEQUENCE:0
DTEND:20060117T160000
UID:2009-11-17T20:55:58-08:00_72732409@limen.stat.ucla.edu
DESCRIPTION:P-value quantifies the strength of statistical evidence against
  a null/uninteresting/default hypothesis. In some data-abundance areas\, re
 searchers often compute an array of p-values for multiple decision making. 
 For designing a follow-up study\, ordered p-values can guide the prioritiza
 tion of the problem list. Two issues arise: (1) How to estimate the number 
 of false positives W(k) in the first k problems with the smallest p-values\
 ; (2) How many problems should we recommend in order to include at least u*
  true positives in the follow-up list with high probability? To address the
  first issue\, we first convert p-value from p to Y=-log(1-p). We then deri
 ve a simple formula\, cAMP(k)\, to estimate W(k) by exploiting the memoryle
 ss property of the exponential distribution. Using the martingale theory\, 
 we show cAMP(T) gives a conservative estimate of W(T-1) for any step-down F
 DR (false discovery rate) procedure that rejects the first T-1 null hypothe
 ses. We also show how to use the cAMP sequence for addressing the second is
 sue.
SUMMARY:Extreme P-values\, False Positives\, and a Theory of Memoryless Con
 version
DTSTART:20060117T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:108
SEQUENCE:0
DTEND:20060112T160000
UID:2009-11-17T20:55:58-08:00_113831344@limen.stat.ucla.edu
DESCRIPTION:Causal inference is best understood using potential outcomes\, 
 which include all post treatment quantities. The use of potential outcomes 
 to define causal effects is particularly important in more complex settings
 \, i.e.\, observational studies or randomized experiments with complication
 s such as noncompliance. Here we deal with the issue of estimating the casu
 al effect of a treatment on a primary outcome that is censored by an interm
 ediate outcome\, for example\, the effect of a drug treatment on Quality of
  Life (QOL) in a randomized experiment where some of the patients die befor
 e their QOL can be assessed.  Because both QOL and death are post-randomiza
 tion quantities\, they both should be considered potential outcomes\, and t
 he effect of treatment versus control on QOL is only well-defined for the s
 ubset of patients who would live under either treatment or control. Another
  application is to an educational program designed to increase final test s
 cores\, which are not defined for those who drop out of school before takin
 g the test. A further application is to studies of the effect of job traini
 ng programs on wages\, where wages are only defined for those who are emplo
 yed\, and thus the effect of the job-training program on wages is only well
 -defined for the subset of individuals who would be employed whether or not
  they were trained. Some empirical results are presented from Zhang\, Rubin
  and Mealli (2004)\, which indicate that this framework can lead to new ins
 ights because the analysis is not predicated on traditional econometric ass
 umptions. \n\n Note that this is a joint seminar with Bio-stat\, on a Thurs
 day and special location 53-105 CHS.
SUMMARY:Causal Inference Through Potential Outcomes: Application to Quality
  of Studies with 'Censoring' Due to Death and to Studies of the Effect of J
 ob-training Programs on Wages
DTSTART:20060112T150000
DTSTAMP:20091117T205558
LOCATION:53-105 Center for Health Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:111
SEQUENCE:0
DTEND:20051129T160000
UID:2009-11-17T20:55:58-08:00_60191036@limen.stat.ucla.edu
DESCRIPTION:Sparse and  moderately strong signals are naturally found in ma
 ny applications\, e.g.  covert communications\,  Kuiper Belt Objects.   Her
 e\, sparsity refers to the situation that out of a massive data set\,  only
  a small fraction of data points contain relevant information or signal\, w
 hile others are irrelevant or noise\;  moderately strong signals are those 
 stronger than a typical noise\, but not stronger than all of them. \n\n Spa
 rse inference in the regime of moderate significances poses interesting new
  phenomenon as well as new  challenges.   In particular\,  the signals are 
 not strong enough to stand out for themselves\, and it is not possible to i
 ndividually tell which is  signal and which is noise.    However\,  valid i
 nference is still possible\, and recently there has been a lot of interests
  in the  proportion} of the signals. \n\n In this talk\, we introduce three
  recent inference tools for the proportion of signals:  the Higher Criticis
 m\, Meinshausen and Rice's confidence lower bound\, and Cai\, Jin\, and Low
 's confidence lower bound.   We will discuss the asymptotic behavior of the
 se tools in detail\, and also compare their strength and weakness.
SUMMARY:New Tools for Sparse Inference: The Regime of Moderate Significance
 s
DTSTART:20051129T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:112
SEQUENCE:0
DTEND:20051122T160000
UID:2009-11-17T20:55:58-08:00_576635215@limen.stat.ucla.edu
DESCRIPTION:Experiments with functional responses are becoming increasingly
  common in industrial and engineering studies.  We consider analysis of exp
 eriments with functional linear models y(t) = X b(t) + e(t)\, where y(t) is
  a response function vector\, X is a fixed covariate matrix\, b(t) is an un
 known coefficient function vector and e(t) is an independent Gaussian error
  function vector. The F test proposed by Shen and Faraway (Statistica Sinic
 a 2004) is used to test nested linear models. Chi-square quantile-quantile 
 plots are proposed to check the assumption of Gaussian error process and ou
 tliers.  Jackknife residuals and an associated test are proposed to detect 
 outliers. Cook's distance is  defined to detect influential cases. The meth
 odology is illustrated by an example from a robust design study. This is jo
 int work with Qing Shen.
SUMMARY:Analysis of Experiments With Functional Responses
DTSTART:20051122T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:113
SEQUENCE:0
DTEND:20051117T170000
UID:2009-11-17T20:55:58-08:00_224237726@limen.stat.ucla.edu
DESCRIPTION:Since the seminal work of Gelfand and Smith (1990)\, the past 1
 5 years have witnessed a host of techniques designed to improve the speed o
 f the Gibbs sampler and more generally Markov chain Monte Carlo (MCMC) in r
 outine applications.  Re-parametrizations\, or variable transformations\, a
 re known to be a key to efficient implementation.  However\, the majority o
 f proposals to date focus on either a single transformation or on straightf
 orward combinations of several transformations\, as with Meng's and van Dyk
 's (1997\, 1999\, 2001) efficient Data Augmentation (DA) strategy.  Here we
  demonstrate that by interweaving two specific kinds of (one-to-one) DAs or
  transformations\, we can gain considerable speed in convergence and simpli
 city in construction.  Simplicity is due to choosing transformations via su
 fficiency and ancillarity\, two familiar classical concepts.  In addition\,
  by using conditional sufficiency and ancillarity\, we can interweave diffe
 rent transformations for different steps (e.g.\, Gibbs steps) within each i
 teration\, and thereby the proposed strategy provides a fairly general reci
 pe for constructing a new generation of efficient algorithms for complicate
 d applications\, such as generalized linear mixed models.  To demonstrate t
 his empirically\, we conduct posterior simulation under a parameter driven 
 Poisson time series model (Cox\, 1981) fitted to a Chandra X-ray data set. 
  A historical polio incidence data set (Zeger\, 1988) and simulated data ar
 e also used for empirical demonstration.  Theoretically\, we show that\, in
  addition to being robust\, under certain conditions this interweaving stra
 tegy leads to the same optimal implementation among a broad class of DA sch
 emes\, as formulated by Liu and Wu (1999).  These findings suggest a host o
 f open questions\, including a full explanation of why interweaving suffici
 ency and ancillarity can lead to such successes for MCMC...
SUMMARY:Espousing Modern Computation with Classical Statistics:  Sufficienc
 y\, Ancillarity and A New Generation of MCMC
DTSTART:20051117T160000
DTSTAMP:20091117T205558
LOCATION:6629 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:117
SEQUENCE:0
DTEND:20051011T160000
UID:2009-11-17T20:55:58-08:00_904806885@limen.stat.ucla.edu
DESCRIPTION:Curves play an important role in statistical analysis of shapes
 . For instance\, objects in 2D images can be characterized by shapes of the
 ir boundaries\, or shapes of surfaces of 3D objects can be studied through 
 shapes of certain level curves on these surfaces. A fundamental tool in ana
 lyzing shapes of closed curves is the construction of geodesics between any
  two such curves. In past we have used a shooting method for constructing g
 eodesics on spaces of planar (R^2)\, closed curves and have studied the res
 ulting shape statistics. This talk will have two parts: In the first part\,
  I describe the basic ideas and applications from such statistical analysis
  of shapes. In the second part\, I present an extension that uses a path-st
 raightening approach for finding geodesics between closed curves in $R^3$. 
 The basic idea is to connect the given two curves on an appropriate manifol
 d using any path\, and to iteratively straighten this path until it becomes
  a geodesic. I will illustrate this approach using examples from computer v
 ision. \n\n (This work is performed in collaboration with Eric Klassen).
SUMMARY:A Statistical Analysis on Shape Spaces of Closed Curves
DTSTART:20051011T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:119
SEQUENCE:0
DTEND:20051004T160000
UID:2009-11-17T20:55:58-08:00_155340168@limen.stat.ucla.edu
DESCRIPTION:ChIP-chip (or ChIP-on-chip) is a technique for isolation and id
 entification of the DNA fragments that are occupied by specific DNA binding
  proteins. The ChIP-chip data can be obtained over the whole genome in the 
 form of a one-dimensional series of signals\, where a peak is generally pre
 sent at a protein binding site. In this talk\, I will give a description of
  the ChIP-chip data\, and present a probability model for such data. I will
  also describe a software called Mpeak that we developed for identifying an
 d testing the peaks for locating the potential binding sites. The talk is b
 ased on joint work with Y. Wu and Ren's lab in UCSD. \n\n This work has bee
 n published in Nature.
SUMMARY:A Probability Model of ChIP-chip Data
DTSTART:20051004T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:115
SEQUENCE:0
DTEND:20051025T160000
UID:2009-11-17T20:55:58-08:00_133246147@limen.stat.ucla.edu
DESCRIPTION:To understand gene interactions and then a more complete portra
 yal of gene dynamics\, researchers usually put together the gene expression
  datasets from multiple-experimental conditions to identify functionally re
 lated genes. In this talk\, I will discuss some statistical issues concerni
 ng such studies\, and introduce several non-standard methods we have develo
 ped for finding functionally related genes. In particular\, I will introduc
 e a clustering analysis of Serial Analysis of Gene Expression (SAGE) data u
 sing Poisson assumptions\, a model for capturing gene associations across m
 ultiple dependent experimental conditions\, etc. Our methods have been foun
 d to be advantageous in applications. \n\n Bio sketch:<br/> 1997-2001 PhD A
 pplied Mathematics (Computational Biology)\, USC advisor: Larry Goldstein<b
 r/> 2001-2003 Post-doc (computational Biology)\, Harvard University supervi
 sors: Wing Wong and Jun Liu<br/> 2003-present Assistant Professor\, Departm
 ent of Statistics\, UC Berkeley<br/>
SUMMARY:Non-standard Methods for Finding Functionally Related Genes using G
 ene Expression Data
DTSTART:20051025T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:121
SEQUENCE:0
DTEND:20050531T160000
UID:2009-11-17T20:55:58-08:00_42505049@limen.stat.ucla.edu
DESCRIPTION:In many fields of biomedical research\, especially those on chr
 onic diseases\, missing data are commonly met in studies with longitudinal 
 designs.  Without proper handling of missing values\, either biased estimat
 ion or invalid inferences would be obtained in statistical longitudinal dat
 a analysis.  For longitudinal data with high dimensions\, we usually observ
 e two types of missingness: intermittent missing and dropout.  Intermittent
  missing values are usually due to participants occasionally missing of cli
 nic visits in clinical trials or nonresponses in survey studies.  Dropout m
 issing refers to participants premature withdrawal from the study.  In many
  practical settings\, the assumption of ignorability is acceptable for the 
 intermittent missingness but not for dropout.  In this talk\, I will first 
 define the concept of ignorability by classifying missingness mechanisms in
 to three groups: outcome-dependent missingness\, pattern-dependent missingn
 ess\, and shared-parameter missingness.  Correspondingly\, three modeling a
 pproaches will be introduced to handle three types of nonignorable dropout\
 , i.e.\, selection models\, pattern-mixture models\, and random-effects Mar
 kov transition models.  Then\, a strategy called Multiple Partial Imputatio
 n (MPI) is to be proposed for jointly dealing with ignorable intermittent m
 issing values and nonignorable dropouts. Within the framework of MPI\, inte
 rmittent missing values are first imputed multiple times\; then partially i
 mputed data sets are analyzed using any of the three modeling approaches fo
 r nonignorable dropouts\; and finally multiple analysis results are consoli
 dated to make one inferential statement.  The method will be illustrated by
  using several practical data sets from substances abuse clinical trials.  
 If time permits\, related issues would also be discussed\, e.g.\, strategie
 s for missing-data exploration and assessment\, Bayesian models selection\,
  and functional regression modeling.
SUMMARY:Strategies for Longitudinal Data Analysis with Nonignorable Missing
  Values
DTSTART:20050531T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:122
SEQUENCE:0
DTEND:20050525T160000
UID:2009-11-17T20:55:58-08:00_395033746@limen.stat.ucla.edu
DESCRIPTION:We introduce the area of remote device fingerprinting\, or fing
 erprinting a host (physical device and OS) as opposed to an operating syste
 m or class of devices\, remotely\, and without the fingerprinted device's k
 nown cooperation.  We accomplish this goal by exploiting small\, microscopi
 c deviations in device clock frequency (clock skews).  Our techniques do no
 t require any modification to the fingerprinted devices.  Our techniques re
 port consistent measurements when the measurer is thousands of miles\, mult
 iple hops\, and tens of milliseconds away from the fingerprinted device\, a
 nd when the fingerprinted device is connected to the Internet from differen
 t locations and via different access technologies.  Further\, one can apply
  our passive and semi-passive techniques when the fingerprinted device is b
 ehind a NAT or firewall\, and also when the device's system time is maintai
 ned via NTP or SNTP.  One can use our techniques to obtain information abou
 t whether two devices on the Internet\, possibly shifted in time or IP addr
 esses\, are actually the same physical device.  Example applications includ
 e: computer forensics\; tracking\, with some probability\, a physical devic
 e as it connects to the Internet from different public access points\; coun
 ting the number of devices behind a NAT even when the devices use constant 
 or random IP IDs\; remotely probing a block of addresses to determine if th
 e addresses correspond to virtual hosts\, e.g.\, as part of a virtual honey
 net\; and unanonymizing anonymized network traces.
SUMMARY:Remote Device Fingerprinting
DTSTART:20050525T150000
DTSTAMP:20091117T205558
LOCATION:6704 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:123
SEQUENCE:0
DTEND:20050517T160000
UID:2009-11-17T20:55:58-08:00_494196897@limen.stat.ucla.edu
DESCRIPTION:This will be a very chatty seminar. I would like to discuss som
 e of the questions that arise in  the search for gene responsible for quant
 itative traits. These should be\, as far as I am concerned\, part of an eth
 ical and political debate\, but statistics does have a role to play. Galton
  and Fisher\, at the very beginning of the discipline that we now call Stat
 istics\, were deeply involved in the debate on the genetic origin of comple
 x\, quantitative traits. I will showcase some of the traits that are curren
 tly being investigated and  point to some challenges for statistics.
SUMMARY:Genetics of Quantitative Traits: Looking Forward Standing on the Sh
 oulders of Giants
DTSTART:20050517T150000
DTSTAMP:20091117T205558
LOCATION:6629 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:126
SEQUENCE:0
DTEND:20050426T160000
UID:2009-11-17T20:55:58-08:00_592707676@limen.stat.ucla.edu
DESCRIPTION:Rather than present a specific piece of new research\, I think 
 it would be more informative if I  talk generally about latent variable (LV
 ) models.  The key parameters &theta\; of a wide class of such  models are 
 parameters that generate the population means &mu\; = &mu\;(&theta\;)  and 
 covariances &Sigma\; = &Sigma(&theta\;).
SUMMARY:Linear Latent Variable Models and Related Statistical Issues
DTSTART:20050426T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:127
SEQUENCE:0
DTEND:20050419T160000
UID:2009-11-17T20:55:58-08:00_64899518@limen.stat.ucla.edu
DESCRIPTION:Over the last decade\, the Internet has been described as a dem
 ocratizing force\; a global distribution system providing instantaneous acc
 ess to information and enabling new forms of collaboration. In short\, the 
 Internet connects people and a wide range of information sources and servic
 es. In this talk\, I will consider a new trend in networking in which embed
 ded information technologies create connections between people and the phys
 ical world. I will focus mainly on distributed networks of sensors and actu
 ators\, and various infomechanical devices\, embedded in the physical world
  and operating semi-autonomously. These systems present statisticians and o
 ther data scientists with new methodological challenges\; from design and d
 eployment\, to fault detection and repair\, to visualization\, data analysi
 s and modeling. \n\n Note: This is a dry run of a keynote presentation I am
  supposed to give at the upcoming SIAM conference on data mining. As such\,
  it will be somewhat broad\, uncomfortably speculative\, and at times light
  on formal mathematics. You have been warned.
SUMMARY:Embedded
DTSTART:20050419T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:129
SEQUENCE:0
DTEND:20050405T000000
UID:2009-11-17T20:55:58-08:00_779871854@limen.stat.ucla.edu
DESCRIPTION:We consider the simple linear Boolean model\, an elementary cov
 erage process with applications in biomedicine\, physics\, engineering and 
 other fields.  In the model\, line segments of independent and identically 
 distributed length are located at the points of a Poisson process.  The seg
 ments may overlap\, resulting in a pattern of clumps -- regions of the line
  that are covered by one or more segments -- alternating with uncovered reg
 ions or spacings. We develop an expression for the clump length distributio
 n and a method of successive approximation to solve it.   We show how the m
 ethods can be used to estimate the intensity of the Poisson process and the
  segment length distribution from a sample of clumps and spacings.  We illu
 strate the methods with two applications: particle flow measurement and est
 imation of a treatment effect for a recurrent viral infection.
SUMMARY:The Boolean Model: Estimation and Applications
DTSTART:20050405T000000
DTSTAMP:20091117T205558
LOCATION:C-301 Anderson
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:125
SEQUENCE:0
DTEND:20050503T160000
UID:2009-11-17T20:55:58-08:00_997781052@limen.stat.ucla.edu
DESCRIPTION:Debates have long raged about whether two-dimensional shapes ar
 e inherently defined by boundary curves or by the regions enclosed by those
  boundaries.  This work settles the issue by evaluating two shape models--t
 he boundary curve and the medial axis of the shape--using efficiency of rep
 resentation as the criterion. Considering shapes whose boundaries are simpl
 e\, closed curves with Lipschitz curvature\, we derive a quantitative measu
 re based on medial axis data to determine when the medial axis is a more ef
 ficient shape dsecriptor than the boundary curve. Along the way\, we obtain
  results for the epsilon-entropy of certain compact classes of curves and c
 onstruct adaptive encoding schemes for boundary curves and medial axis data
 .
SUMMARY:Efficient Representation and 2D Shape
DTSTART:20050503T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:131
SEQUENCE:0
DTEND:20050301T160000
UID:2009-11-17T20:55:58-08:00_684283345@limen.stat.ucla.edu
DESCRIPTION:There has been considerable interest recently in collecting and
  analyzing internet traffic data in order to estimate quality of service pa
 rameters such as packet loss rates and delay distributions. Internet servic
 e providers use this information in order to estimate\, provision for\, and
  monitor network service quality. There are a number of interesting statist
 ical inverse problems that arise in this area\, which has been termed netwo
 rk tomography. This talk will give an overview of these problems and focus 
 on the so-called active network tomography problem.  The goal here is to es
 timate network losses and delay distributions by actively probing the netwo
 rk. We describe several (hopefully) interesting results on the design of pr
 obing experiments and estimation methods. Some applications of the results 
 to network monitoring will also be illustrated. This is joint work with Geo
 rge Michailidis\, Bowei Xi\, and Earl Lawrence.
SUMMARY:Some Statistical Issues in Network Tomography
DTSTART:20050301T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:132
SEQUENCE:0
DTEND:20050222T160000
UID:2009-11-17T20:55:58-08:00_341237553@limen.stat.ucla.edu
DESCRIPTION:(joint work with Robert Anderson) \n\n An effect display is a g
 raphical or tabular summary of a statistical model based on high-order term
 s in the model. Effect displays have previously been defined by Fox (1987\,
  2003) for generalized linear models (including linear models). After revie
 wing and illustrating effect displays for generalized linear models\, we ex
 tend the displays to two models commonly used for polytomous categorical da
 ta: the multinomial logit model and the proportional-odds logit model. Whil
 e most details of effect displays for these two models are straightforward\
 , the derivation of standard errors is more challenging. We provide formula
 s for standard errors and develop examples.
SUMMARY:Effect Displays for Multinomial and Proportional-Odds Logit Models
DTSTART:20050222T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:133
SEQUENCE:0
DTEND:20050217T160000
UID:2009-11-17T20:55:58-08:00_394192174@limen.stat.ucla.edu
DESCRIPTION:One goal of computational biology is to find stretches of DNA s
 equence that are mutating at unusually slow rates. Such conserved elements 
 often turn out to have important functional roles in the cell. Comparing se
 quence data in specific regions across different organisms has been effecti
 ve in detecting conserved elements. I will describe the phylogenetic shadow
 ing approach to genome comparison\, which uses close evolutionary relatives
  like humans and some primate species. I will present results on detecting 
 regulatory elements such as transcription-factor binding motifs. Then I wil
 l discuss a different problem\, the identification of conserved genes\, usi
 ng a hidden Markov model which incorporates phylogeny. Power analysis can b
 e used to suggest species that are worth sequencing\, in order to identify 
 conserved regions. I will talk about the underlying assumptions and the emp
 irical results.
SUMMARY:Statistical Methods for Genome Comparison
DTSTART:20050217T150000
DTSTAMP:20091117T205558
LOCATION:5422 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:136
SEQUENCE:0
DTEND:20050201T160000
UID:2009-11-17T20:55:58-08:00_731720256@limen.stat.ucla.edu
DESCRIPTION:Asset returns in empirical and theoretical literature are often
  quadratic returns: the short rate\, the maximal squared Sharpe ratio\, the
  hedging covariance vector\, and unspanned covariance matrix are all quadra
 tic functions of quadratic processes (the drift and diffusion of a quadrati
 c process are quadratic functions of the process itself). In this paper\, I
  explicitly solve dynamic portfolio choice problems\, up to the solution of
  an ordinary differential equation (ODE)\, when the asset returns are quadr
 atic and the agent has a constant relative risk aversion coefficient. My so
 lution includes as special cases most existing explicit solutions of dynami
 c portfolio choice problems. I also present applications that are not in th
 e literature\, such as the bond portfolio selection problem when the bond r
 eturns are described by quadratic term structure models\, the stock portfol
 io selection problem when the stock return volatility is stochastic as in H
 eston (1993)\, and a bond and stock portfolio selection problem when bond r
 eturns are described by Cox\, Ingersoll\, and Ross\' (1985) (CIR) model and
  stock returns display stochastic volatility. The stochastic interest rate{
 stochastic volatility model is calibrated to the US market data\; dynamic c
 hoice effects on the bond are strong (weak) when investment horizon is 5 ye
 ars (1 month) and dynamic choice effects on the stock is weak at the calibr
 ated parameters. Furthermore\, the ratio of bond over stock portfolio weigh
 ts increases with risk aversion when the investment horizon is 5 years\, th
 us providing a potential resolution to an asset allocation puzzle of Canner
 \, Mankiw\, and Weil (1997). In contrast to static models\, I demonstrate t
 he following properties of dynamic portfolio weights: a risk averse agent m
 ay short a risky asset with a strictly positive risk premium and a more ris
 k averse agent may hold more risky assets with a strictly positive risk pre
 mium.
SUMMARY:Portfolio Selection in Stochastic Environments
DTSTART:20050201T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:137
SEQUENCE:0
DTEND:20050125T160000
UID:2009-11-17T20:55:58-08:00_735955610@limen.stat.ucla.edu
DESCRIPTION:Environmental and Speaker Robustness are important for the depo
 lyment of speech recognition systems in the real world. This talk will addr
 ess the issues from a statistical perspective. Environmental robustness is 
 discussed in terms of a ploynomial regression of utterance signal-to-noise 
 ratio to deal with non-stationary background noise.  To cope with speaker v
 ariation\, structured maximum likelihood linear regression is investigated 
 and the transformations are obtained as the weighted average of competing s
 tructures based on the minimum description length.
SUMMARY:Towards Environmental and Speaker Robustness in Automatic Speech Re
 cognition
DTSTART:20050125T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:138
SEQUENCE:0
DTEND:20050118T160000
UID:2009-11-17T20:55:58-08:00_24058195@limen.stat.ucla.edu
DESCRIPTION:This talk is devoted to the decomposition of an image $f$ into 
 a sum of two components $u+v$\, where $u$ is a function of bounded variatio
 n (a cartoon approximation of $f$)\, while $v$ is an oscillating function\,
  representing texture or noise. To model the component $v$\, we use a space
  of oscillatory functions\, defined by duality\, instead of the standard $L
 ^2$ norm. The obtained algorithm is very simple\, making use of differentia
 l equations and is easily solved in practice. Finally\, I will present vari
 ous numerical results on real images\, showing the obtained decomposition $
 u+v$. I will also illustrate how the proposed method can be used for image 
 restoration and segmentation.
SUMMARY:Image Decomposition and Function Spaces
DTSTART:20050118T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:139
SEQUENCE:0
DTEND:20050114T160000
UID:2009-11-17T20:55:58-08:00_701498199@limen.stat.ucla.edu
DESCRIPTION:"We propose spatial-temporal models for sequences of images of 
 dynamic scenes that exhibit some temporal regularity properties\, intended 
 in a statistical sense\; these include\, for example\, ocean waves\, smoke\
 , whirlwind\, fire\, foliage\, but also moving objects with a defined shape
 \, for instance flowers\, or flags in wind etc. We call such motion scenes 
 as dynamic textures. This work presents a characterization of dynamic textu
 res\, and poses the problems of modeling\, learning\, synthesis\, animation
 \, recognition\, and segmentation of dynamic textures. \n\n Since from imag
 es alone (i.e. in absence of any additional prior knowledge) the problem of
  inferring the physical model that could have generated them is ill-posed\,
  in this work we analyze sequences of images solely as visual signals. We d
 o so by building a statistical framework\, and draw on disciplines like tim
 e series analysis\, system\, control\, and identification theory. \n\n We d
 erive three generative models\, the simplest possible\, that are able to ca
 pture\, respectively\, the temporal second-order statistics\, the spatio-te
 mporal second-order statistics\, and the higher-order temporal statistics o
 f dynamic textures. We propose to learn model parameters in the maximum-lik
 elihood sense\, or minimum prediction error variance. We derive efficient c
 losed-form inference procedures for learning the second-order statistics\, 
 and revert to non-linear optimization techniques for the higher-order ones.
  After learning a model\, it can be used to extrapolate\, or predict new im
 age data both in the temporal and spatial domain. We analyze the meaning of
  the parameters of a model\, and show how they can be manipulated to contro
 l\, or animate the simulation. Using the geometry of subspaces\, and statis
 tical pattern recognition theory we derive a technique to discriminate betw
 een models\, and assess the potential for building a recognition system. Fi
 nally\, by combining these results with a variational framework\, we design
  a region-based segmentation system able to partition a video sequence into
  regions characterized by different spatio-temporal statistics."
SUMMARY:Statistical Modeling of Complex Motion&mdash\;Ocean Waves\, Smoke\,
  Whirlwind\, Fire\, etc.
DTSTART:20050114T150000
DTSTAMP:20091117T205558
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:140
SEQUENCE:0
DTEND:20050111T160000
UID:2009-11-17T20:55:58-08:00_814941397@limen.stat.ucla.edu
DESCRIPTION:Collaboration will be as much at the heart of solving complex o
 r multi-disciplinary problems for the next 25 years as the interface of com
 puting and statistics has been for the past 25 years.  A five-stage process
  for solving these problems completely is characterized. Then the critical 
 success factors of collaboration are integrated into this problem-solving p
 rocess. \n\n Of the papers read and presentations heard in 2004: 1/3 solved
  problems incompletely\, 1/3 collaborated insufficiently\, and only 1/3 exh
 ibited good science. That most of them were in proteomics\, the current fro
 ntier in life-science research\, is most disturbing. \n\n Checklists of wha
 t to do and how to do it are described in detail\, and scorecards to assess
  how well it is done are discussed.  The most complex things are simpler wi
 th checklists and scorecards.
SUMMARY:Collaborative Team's Guide to Solve Problems by Maximizing the Valu
 e to a Client when Problems and Data are Complex or Multi-Disciplinary
DTSTART:20050111T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:145
SEQUENCE:0
DTEND:20041102T160000
UID:2009-11-17T20:55:58-08:00_910163933@limen.stat.ucla.edu
DESCRIPTION:We outline a unified approach to setting up a model for either 
 the covariance or correlation matrix of repeated measures and discuss MCMC 
 computational methods for this approach. The model bridges two important ap
 proaches: (i) the Inverse-Wishart prior distribution with given scale matri
 x\, and (ii) parametric models such as those estimated in SAS Mixed and S-p
 lus lme.  Applications to data from the UCLA Brain Injury Research Center w
 ill be discussed.
SUMMARY:Modeling the Dependence Stucture of Repeated Measures Data
DTSTART:20041102T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:143
SEQUENCE:0
DTEND:20041116T160000
UID:2009-11-17T20:55:58-08:00_623097522@limen.stat.ucla.edu
DESCRIPTION:Automatic image orientation detection for photos is a useful\, 
 yet challenging research topic. Humans use scene context and semantic objec
 t recognition to identify the correct image orientation. However\, it is di
 fficult for a computer to perform the task in the same way because current 
 object recognition algorithms are extremely limited in their scope and robu
 stness. \n\n First\, we highlight the findings of a psychophysical study co
 nducted to understand human perception of image orientation. The accuracy b
 y humans provides an upper bound for the performance of an automatic system
 . In addition\, the use of a large\, representative image set (photo space)
  and extensive interaction with the observers reveal cues used by humans at
  various resolutions. Finally\, we present a probabilistic approach to imag
 e orientation detection by integrating low-level and selected semantic feat
 ures. Our current accuracy is 90%\, promising given the findings of the psy
 chophysical study.
SUMMARY:Image Understanding for Photos: How do Humans and Computers Recogni
 ze Image Orientation?
DTSTART:20041116T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:141
SEQUENCE:0
DTEND:20041207T160000
UID:2009-11-17T20:55:58-08:00_497878688@limen.stat.ucla.edu
DESCRIPTION:Multiscale image analysis has recently led to theory in describ
 ing the fundamental difficulty and intrinsic computational complexity of im
 age detection tasks. I will review some of these results. During these rese
 arch\, new statistical problems arised. For example\, the limit distributio
 n of the length of a longest significant run in a random net needs to be de
 rived. (The problem formulation will be described.) I will present our rece
 nt results regarding these statistical problems. This talk will be descript
 ive\, instead of mathematically rigor. Some publications related to this pr
 esentation can be downloaded at http://www.isye.gatech.edu/~xiaoming/public
 ation/ and http://stat.stanford.edu/~donoho/reports.html.
SUMMARY:Multiscale Image Analysis and Related Statistical Problems
DTSTART:20041207T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:142
SEQUENCE:0
DTEND:20041119T160000
UID:2009-11-17T20:55:58-08:00_241306445@limen.stat.ucla.edu
DESCRIPTION:Prediction in Machine Learning  and prediction in statistics ar
 e essentially equivalent fields\, but with different emphasis. I will try t
 o illustrate the relation between theory and practice in this large area by
  a few examples and results.  In particular I will try to address an appare
 nt puzzle: Worst case analyses\, using empirical process theory\, seem to s
 uggest that even for moderate data dimension and reasonable sample sizes go
 od prediction (supervised learning) should be very difficult.  On the other
  hand\, practice seems to indicate that even when the number of dimensions 
 is very much higher than the number of observations\, we can often do very 
 well.  The efficacy of cross validation will also be discussed and some res
 earch directions pointed out.
SUMMARY:Travels on the Frontiers of Statistics and Computer Science
DTSTART:20041119T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:135
SEQUENCE:0
DTEND:20050208T160000
UID:2009-11-17T20:55:58-08:00_803703088@limen.stat.ucla.edu
DESCRIPTION:Recent advances in nano-technology allow scientists for the fir
 st time to follow a biochemical process on a single molecule basis. These a
 dvances also raise many challenging data-analysis problems and call for a s
 ophisticated statistical modeling and inference effort. First\, by zooming 
 in on single molecules\, recent single-molecule experiments revealed that m
 any classical models derived from oversimplified assumptions are no longer 
 valid. Second\, the stochastic nature of the experimental data and the pres
 ence of latent processes much complicate the inference. In this talk we wil
 l use the modeling of subdiffusion phenomenon in enzyme reaction and the in
 ference of DNA hairpin kinetics to illustrate the statistical challenges in
  single-molecule biophysics. \n\n This talk is based on joint works with th
 e Sunney Xie group.
SUMMARY:Stochastic Modeling and Inference in Single Molecule Biophysics
DTSTART:20050208T150000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:116
SEQUENCE:0
DTEND:20051018T160000
UID:2009-11-17T20:55:58-08:00_868250531@limen.stat.ucla.edu
DESCRIPTION:High-throughput expression profiling allows researchers to stud
 y gene activities globally. Genes with similar expression profiles are like
 ly to encode proteins that participate in a common structural complex\, met
 abolic pathway\, or biological process. However\, the converse of this assu
 mption usually does not hold and many biologically related genes do not sho
 w strong correlations. In Li K.C. (2002)\, the author suggested that co-exp
 ression patterns between two proteins in a common biological process may ch
 ange under different "cellular states". Under this assumption\, the direct 
 measurement of association between two biologically related genes are likel
 y to be weakened by the changes of the association levels among different c
 ellular states. In order to find the hidden cellular conditions\, Li propos
 ed the statistics\, liquid association (LA)\, to search for the third gene 
 which might mediate the association changes between two genes. Here\, we pr
 oposed Context-dependent clustering (CDC) to directly find the hidden cellu
 lar conditions from any given two set of biologically related genes. The ap
 plication of CDC on the transcription factor and its binding targets sugges
 ts that the clusters found via CDC is biologically meaningful. The theoreti
 cal aspect and possible generalization of CDC will also be discussed.
SUMMARY:Context-dependent Clustering and its Application on Microarray Data
DTSTART:20051018T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:146
SEQUENCE:0
DTEND:20041026T160000
UID:2009-11-17T20:55:58-08:00_623741465@limen.stat.ucla.edu
DESCRIPTION:There are many areas for statisticians to play a role in comput
 er and network security\, and I will briefly mention a few of them. The tal
 k will focus on a specific problem: profiling users for the purpose of dete
 cting masqueraders. Some random graph techniques to address this problem wi
 ll be discussed in both the network and host contexts.
SUMMARY:Statistical Methods for Network and Computer Security
DTSTART:20041026T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:149
SEQUENCE:0
DTEND:20041005T160000
UID:2009-11-17T20:55:58-08:00_353181071@limen.stat.ucla.edu
DESCRIPTION:Fractional factorial designs are among the most widely used exp
 erimental plans in practice. They are classified into two broad types:  reg
 ular designs and nonregular designs. Regular designs  are defined by some d
 efining relations among the factors\; all other designs are generally refer
 red to as nonregular designs.  Nonregular designs become popular in recent 
 years due to their run size economy and flexibility and due to recent advan
 ces in data analysis and design theory. The main question in fractional fac
 torial designs is how to choose and construct efficient or optimal designs.
  This talk will discuss  data analysis strategy\, optimality criteria such 
 as generalized minimum aberration\, and algorithmic and algebraic construct
 ion methods for nonregular designs.
SUMMARY:Some Recent Advances in Nonregular Fractional Factorial Designs
DTSTART:20041005T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:151
SEQUENCE:0
DTEND:20040521T160000
UID:2009-11-17T20:55:58-08:00_274750755@limen.stat.ucla.edu
DESCRIPTION:Graphics are a good means for exploring and presenting data and
  for evaluating models.  Statistical models are valuable for testing ideas 
 and for estimating relationships between variables.  The two complementary 
 approaches need to be integrated more strongly\, especially in the analysis
  of large data sets.  The issues will be illustrated using interactive soft
 ware from the Augsburg group and a large sociological data set discussed in
  the book "Bowling Alone" by Robert Putnam.
SUMMARY:Blending Statistics and Graphics in Visual Data Mining
DTSTART:20040521T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:153
SEQUENCE:0
DTEND:20040427T160000
UID:2009-11-17T20:55:58-08:00_655134494@limen.stat.ucla.edu
DESCRIPTION:In this talk\, we consider the problem of detecting and recover
 ing chirps from noisy data.  Chirps are signals which are neither smoothly 
 varying nor stationary but rather\, which exhibit rapid oscillations and ra
 pid changes in their frequency content. This behavior is very different tha
 n that assumed in the standard literature which typically assumes smoothnes
 s and homogeneity. One particular application of note in conjunction with t
 his line of research is the detection of gravitational waves. \n\n Building
  on recent advances in computational harmonic analysis\, we design librarie
 s of multiscale chirplets\, and introduce detection strategies which are mo
 re sensitive than existing feature detectors. The idea is to use structured
  algorithms which exploit information in the chirplet dictionary to chain c
 hirplets together adaptively as to form chirps with polygonal instantaneous
  frequency\; these structured algorithms are so sensitive that they allow t
 o detect signals whenever their strength makes them detectable by any metho
 d\, no matter how intractable. Formally\, we propose a test statistic which
  provably attains near-optimal decision bounds over a wide range of meaning
 ful classes of chirps. In addition\, there is a way to invoke dynamic progr
 amming and network flows algorithms to rapidly compute our test statistics.
  Similar strategies and results extend to the estimation problem. We hope t
 o report on early numerical experiments. \n\n Parts of this work are joint 
 with Hannes Helgason.
SUMMARY:Chirplets: Multiscale Detection and Recovery of Chirps
DTSTART:20040427T150000
DTSTAMP:20091117T205558
LOCATION:3656 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:155
SEQUENCE:0
DTEND:20040413T160000
UID:2009-11-17T20:55:58-08:00_255058403@limen.stat.ucla.edu
DESCRIPTION:The optimal detection procedure for detecting changes in indepe
 ndent and identically distributed (i.i.d.) sequences in a Bayesian setting 
 was derived by Shiryaev in the nineteen sixties. However\, the analysis of 
 the performance of this procedure in terms of the average detection delay a
 nd false alarm probability has been an open problem. In this work\, we deve
 lop a general asymptotic change-point detection theory that is not limited 
 to a restrictive i.i.d. assumption. In particular\, we investigate the perf
 ormance of the Shiryaev procedure for general discrete-time stochastic mode
 ls in the asymptotic setting where the false alarm probability approaches z
 ero. We show that the Shiryaev procedure is asymptotically optimal in the g
 eneral non-i.i.d. case under mild conditions. We also show that the two pop
 ular non-Bayesian detection procedures\, namely the Page and the Shiryaev-R
 oberts-Pollak procedures\, are generally not optimal (even asymptotically) 
 under the Bayesian criterion. The results of this study are shown to be esp
 ecially important in studying the asymptotics of decentralized change detec
 tion procedures. Several applications such as network security (rapid detec
 tion of attacks in computer networks) and missile defense (target detection
 /tracking) will be addressed.
SUMMARY:A General Change-Point Detection Theory
DTSTART:20040413T150000
DTSTAMP:20091117T205558
LOCATION:3656 Geology Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:148
SEQUENCE:0
DTEND:20041012T160000
UID:2009-11-17T20:55:58-08:00_941342772@limen.stat.ucla.edu
DESCRIPTION:Estimating the size of a population based on capture-recapture 
 data has numerous applications\, including estimating the number of animals
  in wildlife management and fishery\, the number of species in ecological s
 tudies\, the number of patients in epidemiology\, the number of errors in a
  software package\, etc. The case that individuals are allowed to have diff
 erent capture probabilities will be investigated: what we wish to estimate\
 , what we should estimate\, and what we can estimate. The nonidentifiabilit
 y\, instability\, one-sided confidence intervals and construction of lower 
 bounds for the size of a population will be discussed.
SUMMARY:Capture-Recapture\, An Old Song that is Difficult to Sing
DTSTART:20041012T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:110
SEQUENCE:0
DTEND:20051206T160000
UID:2009-11-17T20:55:58-08:00_526857056@limen.stat.ucla.edu
DESCRIPTION:Grammars have been developed and widely studied in language ---
  both natural languages and programming languages. In this talk\, I will di
 scuss grammatical descriptions for natural images\, which are called stocha
 stic graph grammars in 2D images in contrast to the 1D text/speech.  I will
  show that grammars are essential ingredients in representing visual patter
 ns (objects)\, especially for object categories that exhibit large structur
 al variabilities. For example\, funiture\, clothes\, vehicles\, clocks --- 
 also called functional categories in psychology.  From a statistics perspec
 titive\, we are integrating two types of popular models in the literature. 
 The first is stochastoc context free grammar (Markov tree process) for mode
 ling hierarchic structures. The second is Markov random field for represent
 ing context information. In the integrated model\, we mix the Markov random
  fields with SCFG to generate a rich class of models for representing gener
 al visual patterns. That is\, we define a probability on a set of graphical
  configurations which observe both global and local regularities.  By globa
 l regularity\, we mean they are syntactically valid (correct in syntax)\, a
 nd by local regularity\, we mean the local statistical alignments. \n\n Thi
 s is a joint work with David Mumford and a number of students (Hong Chen\, 
 Feng Han\, Jake Porway\, Zijian Xu)
SUMMARY:Mixing Markov Random Fields with Stochastic Graph Grammars
DTSTART:20051206T150000
DTSTAMP:20091117T205558
LOCATION:6629 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:157
SEQUENCE:0
DTEND:20040309T160000
UID:2009-11-17T20:55:58-08:00_972785580@limen.stat.ucla.edu
DESCRIPTION:Research in pharmacogenomics requires massive computer explorat
 ion on heterogeneous databases. The Developmental Therapeutics Program (DTP
 ) at the National Cancer Institute (NCI) launched an unprecedented effort t
 o build up a comprehensive drug sensitivity database about two decades ago.
  In all\, sixty representative human cell-lines from seven cancer types (lu
 ng\, colon\, melanoma\, kidney\, ovary\, brain\, leukaemia) were selected a
 nd their responsiveness over a broad concentration range for tens of thousa
 nds of anti-cancer compounds were tested. The high volume assay has resulte
 d in the identification of many current cancer drugs.  With the project evo
 lving and expanding over years\, more cell-lines and compounds have been te
 sted and the database is continuously updated. On a different front\, the r
 ecent advance of high throughput microarray technology offers a novel molec
 ular characterization of the cell lines. Correlating the drug sensitivity d
 atabase with the gene expression database thus marks a path-breaking post-g
 enome development in the molecular pharmacology of cancer. \n\n However\,  
 it remains puzzling that many drugs of known mechanism turn out uncorrelate
 d with their molecular-target gene expression. We develop a system  for ide
 ntifying candidate genes that intervene\, confound and weaken the drug-gene
  correlation.  Our results link Methotrexate sensitivity to DNA component b
 iosynthesis\, Taxol sensitivity to micro-tubules interacting genes and HIV 
 infection. We also show how the human prion is connected to the expression 
 network of Alzheimer disease. \n\n The key element used in our system is  a
  novel statistical notion called liquid association. It was originally deve
 loped for elucidating the Stanfordís yeast cell-cycle microarray gene expre
 ssion data(Li 2002\, PNAS).  We will discuss why this concept is biological
 ly relevant and how it may play a broader role in refining large scale biol
 ogical data.
SUMMARY:Genomewide Co-expression Dynamics in Yeast and in Human Cell-lines
DTSTART:20040309T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:159
SEQUENCE:0
DTEND:20040224T160000
UID:2009-11-17T20:55:58-08:00_474580813@limen.stat.ucla.edu
DESCRIPTION:This talk will briefly review the dictionary models that have b
 een proposed for non coding regions of DNA and  how their results can be us
 ed to analyze gene expression array experiment. I will focus specifically o
 n our implementation (Vocabulon) and on the analysis of E. Coli gene expres
 sion data that derives from it.
SUMMARY:Dictionary Models for Regulatory Regions in DNA and Gene Expression
  Arrays
DTSTART:20040224T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:161
SEQUENCE:0
DTEND:20040210T160000
UID:2009-11-17T20:55:58-08:00_225587414@limen.stat.ucla.edu
DESCRIPTION:Cognitive Scientists use Graphical Models to model human reason
 ing. These are often contrasted to alternative\, more biologically plausibl
 e\, models such as the Rescorla-Wagner algorithm. We analyze Rescorla-Wagne
 r and show that it can be hand-tuned so that it learns Graphical Models  by
  maximum likelihood. We prove convergence of Rescorla-Wagner for a variety 
 of conditions and estimate its efficiency.
SUMMARY:Graphical Models of Human Reasoning and the Rescorla-Wagner Algorit
 hm
DTSTART:20040210T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:160
SEQUENCE:0
DTEND:20040217T160000
UID:2009-11-17T20:55:58-08:00_771328331@limen.stat.ucla.edu
DESCRIPTION:Visual inference is a daunting task due to the enormous complex
 ity of visual patterns in natural images. The task of inferring and decompo
 sing these visual patterns is called image parsing\, which is one of the ke
 y problems in computer vision and image analysis and has a wide variety of 
 applications such as object recognition\, digital mapping\, medical imaging
 \, stereo matching\, tracking\, etc. \n\n For general and global optimal so
 lutions\, we are forced to use Markov chain Monte Carlo techniques which ma
 ke inference for generative models in the Bayes framework\; but for computa
 tional feasibility\, we are often forced to adopt discriminative and greedy
  methods to get quick\, partial\, or locally optimal solutions.In this talk
 \, I will present a framework called Image Parsing by Data Driven Markov Ch
 ain Monte Carlo (DDMCMC). The key issues about an efficient design will be 
 discussed. The DDMCMC method\, as a computational paradigm\, should combine
  the advantages of both generative and discriminative approaches and theref
 ore is a viable tool for general visual inference. I'll also introduce an a
 lgorithm called "Swendsen-Wang Cuts"\, which generalizes the well-celebrate
 d "Swendsen-Wang" algorithm and makes efficient moves in MCMC for perceptua
 l grouping. I'll show examples of parsing natural images into various stoch
 astic patterns such as mid-level patterns (generic regions and curves)\, ob
 jects (faces and text)\, and 3D structures. Also\, practical applications o
 f applying the framework in video processing will also be demonstrated in t
 he talk.
SUMMARY:Image Parsing by Data-Driven Markov Chain Monte Carlo
DTSTART:20040217T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:162
SEQUENCE:0
DTEND:20040203T160000
UID:2009-11-17T20:55:58-08:00_868906846@limen.stat.ucla.edu
DESCRIPTION:High-dimensional motion (HDM) refers to various complex motions
  with high degrees of freedom. Examples include the articulation of human b
 ody\, the deformation of elastic shapes and the multi-motion of multiple oc
 cluding targets. Visual analysis of these complex motions involves the infe
 rence or estimation of the underlying high-dimensional motion parameters fr
 om video sequences. Such a high-dimensional inverse problem is very challen
 ging in nature\, mainly due to the manifestation of the curse of dimensiona
 lity\, i.e.\, linear increase of the dimensionality incurs dramatic (e.g.\,
  exponential or combinatorial) increase of the complexity. Lacking of a pri
 nciple way to overcome the curse\, most existing methods are either ineffic
 ient or inaccurate. \n\n This talk mainly introduces our approach that may 
 unify the various HDM analysis problems into a statistical framework\, in w
 hich the HDM is represented by a Markov network\, rather than a centralized
  vector. The benefit of such a distributed representation is that the prohi
 bitive HDM inference tasks can be fulfilled in linear time by the collabora
 tions among the distributed but mutually constrained small-scale visual inf
 erence processes\, as revealed by the study based on probabilistic variatio
 nal analysis of this model. Moreover\, we design collaborative statistical 
 sampling algorithms\, called collaborative particle network algorithms\, to
  implement and instantiate this new approach to HDM analysis. This new appr
 oach is expected to be significantly more efficient\, more scalable and fle
 xible\, and more robust. In addition\, I present our recent results on capt
 uring human motion and tracking multiple targets from video. \n\n \n\n \n\n
  Bio: Ying Wu received the Ph.D. degree in electrical and computer engineer
 ing from the University of Illinois at Urbana-Champaign (UIUC) in 2001\, th
 e M.S. degree from Tsinghua University\, China\, in 1997\, and the B.S. deg
 ree from Huazhong University of Science and Technology\, China\, in 1994. H
 e has been an assistant professor at the department of electrical and compu
 ter engineering at Northwestern University since 2001. He was a research in
 tern at Microsoft Research in 1999 and 2000. His current research interests
  include computer vision/graphics\, statistical methods\, image/video under
 standing and multimedia. He received the Robert T. Chien Award at UIUC in 2
 001\, and is a recipient of the NSF CAREER Award.
SUMMARY:Visual Analysis of High-Dimensional Motion
DTSTART:20040203T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:164
SEQUENCE:0
DTEND:20040120T160000
UID:2009-11-17T20:55:58-08:00_447406012@limen.stat.ucla.edu
DESCRIPTION:In this talk I am interested in modeling the relationship betwe
 en a scalar\, Y\, and a functional predictor\, X(t). I introduce a highly f
 lexible approach called Functional Adaptive Model  Estimation (FAME) which 
 extends generalized linear models (GLM)\, generalized additive models (GAM)
  and projection pursuit regression (PPR) to  handle functional predictors. 
 The FAME approach can model any of the  standard exponential family of resp
 onse distributions that are assumed for GLM or GAM while maintaining the fl
 exibility of PPR. For example standard linear or logistic regression with f
 unctional predictors\, as well as  far more complicated models\, can easily
  be applied using this approach. A functional principal components decompos
 ition of the predictor functions is used to aid visualization of the relati
 onship between X(t) and Y. I will also show how the FAME procedure can be e
 xtended to deal with multiple functional and standard finite dimensional pr
 edictors\, possibly with missing data. The FAME approach will be illustrate
 d on  the prediction of five year survival for a patient given observations
  of  blood chemistry levels over time and the prediction of arthritis based
  on bone  shape. I will end with a discussion of the relationships between 
 standard regression approaches\, their extensions to functional data and FA
 ME. \n\n http://www-rcf.usc.edu/~gareth
SUMMARY:Functional Adaptive Model Estimation
DTSTART:20040120T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:165
SEQUENCE:0
DTEND:20040113T160000
UID:2009-11-17T20:55:58-08:00_354211999@limen.stat.ucla.edu
DESCRIPTION:False Discovery Rate (FDR) controlling procedures are multiple 
 testing hypothesis procedures that seek to control the proportion of false 
 positives rejected\, whereas more traditional procedures seek to control th
 e so-called Family Wise Error Rate\, the probability of any false positives
  occuring. The FDR controlling procedures give sensible answers when the te
 sts are independent\, or when the number of tests is very large\, as in mic
 roarray or neuroimaging data. One weakness of the FDR  in problems with spa
 tial structure\, such as neuroimaging\, is the fact that\, for large sample
 s\, the FDR is strictly a marginal measure of error and no use is made of t
 he spatial structure of the data. After reviewing some basics on fMRI (func
 tional magnetic resonance imaging)\, we will review FDR controlling procedu
 res and provide an alternative way of looking at the procedures. With empha
 sis on the application to fMRI we will suggest ways of incorporating spatia
 l dependence into the FDR procedures\, while retaining the simple interpret
 ation of the FDR. Some of the talk is based on joint work with John Storey 
 and David Siegmund\, while other parts are based on joint work with Brian K
 nutson and Keith Worsley.
SUMMARY:Incorporating Spatial Structure in FDR Procedures with Applications
  to Inference in fMRI
DTSTART:20040113T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:166
SEQUENCE:0
DTEND:20031205T160000
UID:2009-11-17T20:55:58-08:00_336096878@limen.stat.ucla.edu
DESCRIPTION:Tree based models are quite popular due to their versatility an
 d interpretability. It is often possible to communicate the results to doma
 in experts with less profound statistical knowledge. On the other hand tree
 -based classifiers can outperform many competing methods if tree models are
  augmented by the use of  whole ensembles of trees or forests. Unfortunatel
 y the interpretability of the individual models is lost in such process. In
  this talk we want to take a closer look at forests with interactive graphi
 cs\, in order to see why and when ensembles work. We will also present some
  methods of exploratory model analysis that can be used to extract addition
 al information about the underlying data structure from the seemingly unpen
 etrable forests\, allowing an assessment of the models in respect to the da
 ta analytic goal and making a compromise between prediction accuracy and in
 terpretation possible.
SUMMARY:Some Light in a Dark Forest: A Closer Look at Tree Model Ensembles
DTSTART:20031205T150000
DTSTAMP:20091117T205558
LOCATION:5225 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:172
SEQUENCE:0
DTEND:20031028T160000
UID:2009-11-17T20:55:58-08:00_196059310@limen.stat.ucla.edu
DESCRIPTION:In neuroscience\, it is often necessary to detect neural activi
 ty that exhibits certain temporal pattern.  This problem can be formulated 
 as detection of patterned clusters of points ("targets").  I will describe 
 an approach to this problem which also applies to detection for continuous-
 valued signals.  Detection is considered classification based on likelihood
  ratio.  Under certain Poisson assumptions on the point processes\, the cla
 ssification is equivalent to linear filtering.  I will present results usin
 g this approach in neuroscience\, such as replayed pattern in spontaneous s
 leep neuronal activity of the birdsong system. Large deviations involved in
  this approach will also be discussed. \n\n Biosketch: Zhiyi Chi graduated 
 from the Division of Applied Mathematics with a PhD in 1998 for work on 'Pr
 obability Models for Complex Systems'. Since then he has been Assistant Pro
 fessor in the Department of Statistics at the University of Chicago.  Besid
 es research in various topics including information theory\, random fields 
 and large deviations\, he has been working with neuroscientists on coding i
 n neural systems.  He is currently a member of the Committee on Computation
 al Neuroscience at the University of Chicago.
SUMMARY:Filtering for Point Processes and Its Applications
DTSTART:20031028T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:169
SEQUENCE:0
DTEND:20031118T150000
UID:2009-11-17T20:55:58-08:00_301652937@limen.stat.ucla.edu
DESCRIPTION:A calibration scale for p-values under alternatives will be giv
 en\, followed by a proposal for answering the more general question What is
  a weight of evidence against a hypothesis? The answer is in the form of hy
 pothesis estimators\, subject to mean squared error loss\, with special pro
 tection given to the null hypothesis. A family of such hypothesis estimator
 s\, one for each parameter value as null against its two-sided complement\,
  is inverted to obtain an acceptability profile. The merits of such profile
 s over traditional confidence intervals will be argued.  The connection bet
 ween hypothesis estimators to p-values as weights of evidence will be provi
 ded.
SUMMARY:Weights of Evidence Against Hypotheses and Acceptability of  Parame
 ter Values
DTSTART:20031118T140000
DTSTAMP:20091117T205558
LOCATION:365 Kinsey Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:170
SEQUENCE:0
DTEND:20031110T160000
UID:2009-11-17T20:55:58-08:00_382025065@limen.stat.ucla.edu
DESCRIPTION:'Mammography' (screening for breast cancer by x-rays) came to b
 e widely accepted in the 1980s\, when a number of clinical trials demonstra
 ted a substantial reduction in risk. Early detection\, before the disease s
 preads\, permitted therapy that was less invasive and more effective. Quest
 ions that remained were largely about efficacy for younger women\, and opti
 mal frequency for older women.  The consensus was challenged in a series of
  papers by two researchers at the Nordic branch of the Cochrane collaborati
 on\, who concluded that mammography does not save lives: instead\, it expos
 es women to unnecessary surgical procedures.  The basis for the critique tu
 rns out to be simple. Studies that found a benefit from mammography were di
 scounted as being of poor quality.  The remaining negative studies were com
 bined in a meta-analysis. The critique therefore rests on judgments of stud
 y quality. These are based on misreadings of the data and the literature. I
 n our view\, the critique has little merit\, and the prior consensus on mam
 mography was correct. This is joint work with Diana Petitti (Kaiser Permane
 nte) and James Robins (Harvard).
SUMMARY:On the Efficacy of Screening for Breast Cancer
DTSTART:20031110T150000
DTSTAMP:20091117T205558
LOCATION:5200 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:176
SEQUENCE:0
DTEND:20030930T160000
UID:2009-11-17T20:55:58-08:00_254257727@limen.stat.ucla.edu
DESCRIPTION:Treating transcript abundances as quantitative traits and mappi
 ng quantitative trait loci (QTL) for such traits has been the focus of rece
 nt large-scale gene expression studies.  Because quantitative trait mapping
  provides a degree of causal information pertaining to a trait of interest 
 (i.e.\, DNA variations in the QTL lead to variations in the trait)\, and be
 cause microarray-based expression studies often result in the identificatio
 n of patterns of expression associated with complex disease traits\, it is 
 of interest to determine whether we can infer causal relationships between 
 transcript abundances and disease traits using a combination of expression\
 , genotypic\, and disease trait data.   Here I present a statistical proced
 ure for making such inferences via integration of DNA variation and gene ex
 pression data with clinical trait data in a segregating mouse population.  
 This procedure tests whether variations in relative transcript abundances s
 tatistically support a causative or reactive function in the disease state.
   This procedure allows for the objective identification of genes causally 
 associated with disease traits and provides a generalized information measu
 re for the reconstruction of complex gene networks. \n\n Biosketch: Dr. Sch
 adt joined Rosetta Inpharmatics in 1999 after spending 2 years as a researc
 h scientist leading a computational biology group at Roche Bioscience in Pa
 lo Alto\, CA.  Dr. Schadt led initial efforts at Rosetta to experimentally 
 annotate the human genome using gene expression microarrays.  After Rosetta
  was acquired by Merck\, he formed the Research genetics group at Rosetta t
 o more vigorously pursue the study of the genetics of gene expression\, wit
 h target discovery/validation and pharmacogenomic applications to the basic
  and clinical programs at Merck.  Dr. Schadt holds his Doctorate in Biomath
 ematics from UCLA.
SUMMARY:Making Causal Inferences by Combining Transcription\, Genotypic\, a
 nd Clinical Trait Data in Segregating Populations
DTSTART:20030930T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:173
SEQUENCE:0
DTEND:20031021T160000
UID:2009-11-17T20:55:58-08:00_768394636@limen.stat.ucla.edu
DESCRIPTION:In the field of high energy physics\, most statistical analysis
  is performed using computer codes written within the community without the
  direct involvement of academic statisticians.  Typical methods use likelih
 ood ratios or frequentist construction of confidence intervals\, and benefi
 t from expertise in Monte Carlo simulation which is common and well-develop
 ed in our community.  But in the last ten years\, Bayesian-inspired methods
  have become more common\, with resultant controversies that are familiar t
 o statisticians.  At a series of international workshops beginning in 2000\
 , we have begun to make more contact with statisticians who find our proble
 ms interesting.  I will give examples of controversies which remain unresol
 ved\, illustrating with prototype problems which are simple to state\, incl
 uding: confidence intervals for a parameter (such as neutrino mass) which i
 s known to be positive\; confidence intervals for a Poisson mean in the pre
 sence of "background" and nuisance parameters\; goodness of fit for a likel
 ihood-based fit in many dimensions\; and criteria for discovery of a new pa
 rticle such as the Higgs boson. \n\n Bio: Robert Cousins is Professor of Ph
 ysics at UCLA\, having joined the faculty in 1981 after obtaining physics d
 egrees at Princeton (A.B) and Stanford (Ph.D).  He has been a member of col
 laborations performing diverse experiments in high energy physics (elementa
 ry particle physics) at Fermilab (Illinois)\, Brookhaven (Long Island)\, an
 d CERN (Switzerland).  For more information\, see http://www.physics.ucla.e
 du/people/faculty_members/cousins.html
SUMMARY:Statistics Controversies in Experimental Elementary Particle Physic
 s
DTSTART:20031021T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:182
SEQUENCE:0
DTEND:20030429T160000
UID:2009-11-17T20:55:58-08:00_938733464@limen.stat.ucla.edu
DESCRIPTION:Graphs provide a unique data structure for exploring biological
  data. There are many different graphs that can be constructed based on bio
 logic data. These include metabolic pathways\, protein-protein interactions
  as well as co-citation of genes in the scientific literature. In this talk
  I will consider various methods of using graphs and their properties to pe
 rform exploratory data analysis (EDA) on data from a microarray experiment 
 using different graphs based on biological meta-data.
SUMMARY:Graphs and EDA in Computational Biology
DTSTART:20030429T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:184
SEQUENCE:0
DTEND:20030415T160000
UID:2009-11-17T20:55:58-08:00_380595520@limen.stat.ucla.edu
DESCRIPTION:Recombination has been invoked to explain the disparate evoluti
 onary relationships observed for different genes or sequence segments of a 
 single HIV-1 genome. Many methods have been proposed to infer recombination
  events among viral strains. Most of these methods reconstruct the underlyi
 ng evolutionary relationships among the sequences\, conditional on a model 
 for nucleotide evolution and most rely on frequentist inference. These freq
 uentist recombination identification methods often fall into a sequential t
 esting trap.  The most likely parental sequences and crossover points are i
 dentified using the data and then the certainty of recombination is assesse
 d conditional on this identification.  Recently\, reconstructing evolutiona
 ry relationships using Bayesian models has become popular due to the abilit
 y of Bayesian models to handle complex models of evolution. Bayesian models
  are especially attractive in the detection of recombination events because
  they can avoid the trap by allowing for simultaneous inferences about the 
 presence\, number and location of crossover points and the identification o
 f parental sequences.  After briefly presenting the principles for reconstr
 ucting evolutionary histories\, I will present a frequentist method and the
 n a Bayesian multiple change-point model to infer recombination.
SUMMARY:Inferring Recombination among HIV-1 Subtypes
DTSTART:20030415T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:185
SEQUENCE:0
DTEND:20030408T160000
UID:2009-11-17T20:55:58-08:00_433436541@limen.stat.ucla.edu
DESCRIPTION:The model we review of Computational Anatomy (CA) is a Grenande
 r deformable template\, an orbit generated from a template under groups of 
 diffeomorphisms. The metric space structure on the space of anatomical imag
 es is constructed from the geodesic connecting one anatomical configuration
  to another.  The variational problems specifying these metrics along with 
 the Euler equations of motion for the geodesics in the group of diffeomorph
 isms are reviewed.  Metrics that accommodate photometric variation are desc
 ribed extending the anatomical model to incorporate the construction of neo
 plasm.  Various anatomical examples in CA are examined. \n\n Brief Biograph
 y of the speaker: \n\n Michael I. Miller joined The Johns Hopkins Universit
 y in 1998 as a Professor of Biomedical Engineering and Electrical and Compu
 ter Engineering\, where he leads the Center for Imaging Science.  Dr. Mille
 r has authored papers on magnetic resonance imaging\, brain imaging\, image
  understanding and object recognition\, the information theory of imaging s
 cience and computational anatomy. His research and teaching interests are i
 n the areas of pattern theory\, computational linguistics and was one of th
 e original formalizers of the emerging discipline of Computational Anatomy.
   Dr. Miller has authored a book in Random Point Processes with Donald L. S
 nyder. Dr. Miller received his Masters in Electrical and Computer Engineeri
 ng in 1976 and  his PhD in Biomedical Engineering in 1983 from the Johns Ho
 pkins University. \n\n Relevant papers include: \n\n Grenander and Miller\,
  Computational Anatomy:  An Emerging Discipline\, Quarterly of Applied Math
 ematics\, 1998. \n\n Miller and Younes\, Group Actions\, Homeomorphisms\, a
 nd Matching: A General Framework\, IJCV\, 2001. \n\n Miller\, Trouve and Yo
 unes\, On the metrics and Euler-Lagrange Equations of Computational Anatomy
 \, Annual Review of Biomed. Eng\, 2002.
SUMMARY:On the Metrics and Euler-Lagrange Equations of Computational Anatom
 y
DTSTART:20030408T150000
DTSTAMP:20091117T205558
LOCATION:1200 IPAM Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:186
SEQUENCE:0
DTEND:20030401T160000
UID:2009-11-17T20:55:58-08:00_451742386@limen.stat.ucla.edu
DESCRIPTION:We will discuss ways to approximate sums of stationary and ergo
 dic sequences by martingales.  Necessary and sufficient conditions for cond
 itional central limit theorems are given. The idea shed new light on many o
 pen problems and we will apply it to the kernel estimation problem and empi
 rical processes for linear and nonlinear processes.
SUMMARY:On Martingale Approximations
DTSTART:20030401T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:188
SEQUENCE:0
DTEND:20030304T160000
UID:2009-11-17T20:55:58-08:00_435473762@limen.stat.ucla.edu
DESCRIPTION:A common method for investigating possible connections between 
 air pollution and mortality is to compare daily fluctuations of pollutant c
 oncentrations in a community with the corresponding daily fluctuations of c
 ommunity mortality.  A recent EPA draft report on particulate air pollution
  relies heavily on time series studies of this kind.  I will describe the r
 ecent studies and critically discuss several key statistical issues includi
 ng: concomitant weather variation\, the shape of the mortality response to 
 pollutant exposure\, combining information from multiple monitoring sites\,
  and estimating the effect of mortality harvesting.
SUMMARY:Air Pollution and Mortality
DTSTART:20030304T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:190
SEQUENCE:0
DTEND:20030218T160000
UID:2009-11-17T20:55:58-08:00_678521230@limen.stat.ucla.edu
DESCRIPTION:Hastings Coupling is a method of coupling a set of MCMC sampler
 s to enable sharing of information without disturbing the convergence prope
 rties of the individual chains. This information sharing allows algorithms 
 that have improved efficiency\, particularly in problems where mixing is di
 fficult\, such as multi-modal posterior distributions. We illustrate the Ha
 stings Coupling method by presenting a particular case: the Normal Kernel C
 oupler (NKC). \n\n The Normal Kernel Coupler (NKC) uses a normal kernel den
 sity estimator to create an estimate of the unknown target distribution usi
 ng the set of current state vectors. At each iteration\, one component stat
 e vector is updated using the current density estimate. We show that this s
 ample is ergodic (irreducible\, Harris recurrent\, and aperiodic) for any c
 ontinuous distribution on d-dimensional Euclidean space. In addition\, simu
 lation studies show that NKC outperforms standard MCMC methods on a variety
  of unimodal and bimodal problems. \n\n The NKC is implemented as part of H
 YDRA\, a general-purpose library for MCMC which provides a simple and exten
 sible interface for performing MCMC on general problems.
SUMMARY:Efficient and Adaptive MCMC by Coupling Multiple Samplers
DTSTART:20030218T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:192
SEQUENCE:0
DTEND:20030204T160000
UID:2009-11-17T20:55:58-08:00_827653044@limen.stat.ucla.edu
DESCRIPTION:The use of mathematically based computer models for the study o
 f scientific and engineering processes is ubiquitous. The most basic questi
 on in evaluating such a model is: Does the computer model adequately repres
 ent reality? \n\n Statistical methodology for addressing this question will
  be described within the context of test-bed problems. The proposed six-ste
 p strategy deals with major issues associated with a validation process: qu
 antifying the typically multiple sources of error and uncertainty in comput
 er models\, combining multiple sources of information (e.g.\, from field ex
 periments and computer runs)\, calibrating parameters of the computer model
 \, and assessing model predictions in untested situations. A combination of
  spatial and bayesian statistical tools provides the technical apparatus.
SUMMARY:Statistical Validation of Computer Models
DTSTART:20030204T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:193
SEQUENCE:0
DTEND:20030131T160000
UID:2009-11-17T20:55:58-08:00_502019845@limen.stat.ucla.edu
DESCRIPTION:Many of the statisticians of all generations\, in addition to t
 heir research activities\, have worked hard to bring statistical insights a
 nd methods to important human activities. We can continue and extend that t
 radition: Today\, the pervasive use of software to collect data\, at one en
 d of these activities\, and to present proposals and inferences\, at the ot
 her end\, potentially opens new ways to export our techniques.  To do so\, 
 we must make the software in which those techniques are implemented availab
 le to the rest of the world.  This talk presents a new approach to inter-sy
 stem software interfaces designed to help.
SUMMARY:Who Will Use Statistics?  Making Statistical Software Available
DTSTART:20030131T150000
DTSTAMP:20091117T205558
LOCATION:5225 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:194
SEQUENCE:0
DTEND:20030128T160000
UID:2009-11-17T20:55:58-08:00_636225997@limen.stat.ucla.edu
DESCRIPTION:Model-based motif discovery methods\, such as MEME and Gibbs Mo
 tif Sampler\, are important tools for screening DNA regulatory regions to p
 redict binding sites of transcription factors\, proteins involved in transc
 ription.  Although these methods are very successful\, they can be too gene
 ral\, having been developed for both DNA and protein sequences.  They can b
 e distracted by noisy signals in the data that are not characteristic of tr
 ue transcription factor binding sites. We propose a simple extension to the
  underlying model of these methods to improve the prediction of real sites.
  Our method is based on the observation that examples of real sites show po
 sitional trends in the information content (or base conservation).  We assi
 gn prior distributions to the frequency parameters of the model\, penalizin
 g deviations from a specified level of conservation type. Examples with bot
 h simulated and real data show that these changes improve the algorithm's a
 bility to discover motifs with information content patterns typical of real
  binding sites.
SUMMARY:Detecting DNA Regulatory Motifs by Incorporating Positional Trends 
 in Information Content
DTSTART:20030128T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:195
SEQUENCE:0
DTEND:20030121T160000
UID:2009-11-17T20:55:58-08:00_756400864@limen.stat.ucla.edu
DESCRIPTION:Karl Pearson was in many ways the founder of the modern field o
 f statistics.  The beginnings of his interest in the subject are usually at
 tributed to his collaboration with the biologist W. F. R. Weldon and his ad
 miration for the statistical biology of Francis Galton\, and to his eugenic
  commitments.  Here I take up another\, slightly earlier\, source of his in
 volvement with statistics. \n\n The story begins with his role as a Victori
 an advocate of scientific education\, as a replacement for the traditional 
 curriculum based on classics.  Pearson emphasized &quot\;scientific method&
 quot\; as the basis of a modern education.  Meanwhile he was earning his li
 ving as a teacher of mathematics to engineering students at University Coll
 ege London\, students who generally had little mathematical preparation and
  preferred manual training to mathematical study.  His solution\, following
  a Continental tradition\, was to use graphics to teach mathematics.  With 
 typical enthusiasm\, he came to believe by 1890 that graphical statics coul
 d reverse the transformation of mathematics wrought by Descartes. Descartes
  had shown how geometry could be reduced to algebra\; now algebra would be 
 translated into the more practical and visually appealing form of graphical
  geometry.  Pearson took up statistics as part of this program. Although he
  could never dispense with algebraic formulations\, his perspective remaine
 d strongly geometrical and descriptive for years\, and in some ways through
 out his life.  I think we find here one dimension of his disagreements year
 s later with R. A. Fisher.
SUMMARY:Karl Pearson's Utopia of Scientific Education: From Graphical Stati
 cs to Statistical Mathematics
DTSTART:20030121T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:198
SEQUENCE:0
DTEND:20021210T160000
UID:2009-11-17T20:55:58-08:00_766466374@limen.stat.ucla.edu
DESCRIPTION:We examine the network tomography problem whose objective is to
  recover packet loss rates and delays of all the internal links in a networ
 k by using measurements obtained from nodes located on its periphery.  This
  problem is of increasing importance to network administrators due to the f
 act that it is impossible to have access to every single link in a large de
 centralized network. It is also an example of a large scale statistical inv
 erse problem.  A new monitoring scheme\, involving bicast measurements\, is
  introduced and the theoretical properties of several estimators of the par
 ameters of interest (loss rates and delay distributions) investigated. We a
 lso discuss how to design efficient bicast experiments.
SUMMARY:Active Network Tomography through Bicast Experiments
DTSTART:20021210T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:199
SEQUENCE:0
DTEND:20021203T160000
UID:2009-11-17T20:55:58-08:00_502565781@limen.stat.ucla.edu
DESCRIPTION:Questionnaire data often contains categorical data with both bi
 nary outcomes and multiple-ordered answer categories. In practice\, questio
 nnaires are often hampered by nonresponse problems. In order to proceed wit
 h standard analyses\, multivariate normal multiple-imputation implemented i
 n NORM\, SAS 8.2 and Splus 6 is a common procedure to handle nonresponse. H
 owever\, little is known about the robustness of this imputation method whe
 n the data are not normally distributed. \n\n A study was carried out where
  samples were drawn from the California Healthy Kids Survey\, a data set of
  198262 complete records and 8540 incomplete records which served as the po
 pulation. Nonresponse was simulated using a hot-deck procedure using incomp
 lete records. This way\, real patterns of missing data were obtained. Multi
 variate normal imputation was used to get completed data. Imputed binary re
 sponses were rounded in three different ways. Means for continuous variable
 s\, correlations\, an odds ratio\, t-test\, and logistic regression coeffic
 ients were estimated from the multiply imputed data\, and compared to the k
 nown population parameters. It was found that the multivariate normal imput
 ation model has good coverages in most situations\, even when the imputed v
 ariables were binary. Small differences surface between the rounding method
 s used\, however. \n\n A joint work with Thomas R. Belin and Joseph L. Scha
 fer.
SUMMARY:Robustness of a Multivariate Normal Approximation for Imputation of
  Incomplete Categorical Data
DTSTART:20021203T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:201
SEQUENCE:0
DTEND:20021126T160000
UID:2009-11-17T20:55:58-08:00_133635617@limen.stat.ucla.edu
DESCRIPTION:Sparse Mixture Models have important applications in many areas
 \, such as Signal and Image Processing\,  Genomics\, Covert Communication\,
  etc. In my talk\, I will consider the problems of detecting and estimating
  sparse mixtuers. \n\n Detection:  Higher Criticism is a statistic inspired
  by a multiple comparisons concept mentioned in passing by Tukey (1976). We
  are able to show that the resulting  Higher Criticism statistic is effecti
 ve at resolving a very subtle testing problem: testing whether n  normal me
 ans are all zero versus the alternative that a small fraction is nonzero\; 
 the subtlety  of this `sparse normal means' testing problem can be seen fro
 m work of Ingster (1999) and Jin(2002)\, who studied such problems in great
  detail. In their studies\, they identified an interesting range of cases w
 here the small fraction of nonzero means is so small that the alternative h
 ypothesis exhibits little noticeable effect on the distribution on the p-va
 lues either for the bulk of the tests or for the few most highly significan
 t tests. In this range\, when the amplitude of nonzero means is calibrated 
 with the fraction of nonzero means\, the likelihood ratio test for a precis
 ely-specified alternative would still succeed in separating the two hypothe
 ses. We show that the higher criticism is successful throughout the same re
 gion of amplitude vs. sparsity where the likelihood ratio test would succee
 d. Since it does not require a specification of the alternative\, this show
 s that Higher Criticism is in a sense optimally adaptive to unknown sparsit
 y and size of the non-null effects. While our theoretical work is largely a
 symptotic\, we provide simulations in finite samples. We also show Higher C
 riticism works very well over a range of nonGaussian cases. \n\n Estimation
 :  False Discovery Rate (FDR) controlling procedures were suggested as an e
 stimation tool by Abramovich and Benjamini (1995)\, as a purely data-driven
  adaptive thresholding approach.  A central question to statisticians is\, 
  do FDR-controlling procedures connect to any type of "optimality"? Abramov
 ich et al (2000) studied sparse Gaussian data with the Lp-balls and Lq-loss
  frame and pointed out that\, FDR-controlling procedures are asymptotic Min
 imax. \n\n A natural question is that\, is "asymptotic minimax" limited to 
 Gaussian data\,  or what can we conclude for nonGaussian data?   In this ta
 lk we study sparse Poisson Model and  sparse Exponential Model\, which are 
 important models for  nonGaussian data\,  and have applications in many are
 as  such as Astronomy and  Positron Emission Tomography (PET) as well. We s
 how that\,  under natural and sensible frames\,  FDR-controlling procedures
  are asymptotic minimax and adaptive to unknown sparisity for Exponential a
 nd Poisson models. \n\n The talk is based on joint works with David L. Dono
 ho. \n\n References: \n\n [1]  Abramovich\, F.  and   Benjamini\, Y. and   
 Donoho\, D.  and   Johnstone\, I. (2000). Adapting to Unkown Sparsity by Co
 ntrolling the False Discovery Rate\,  Technical Report \, Statistics Depart
 ment\, Stanford University.<br/> [2]  Donoho\, D.  and  Jin\, J. (2002).  H
 igher Criticism for Detecting Sparse Heterogeneous Mixtures. Technical Repo
 rt \, Statistics Department\, Stanford University.<br/> [3]  Jin\, J.  Dete
 ction Boundary for Sparse Mixtures\, To appear.<br/> [4] Asymptotic Minimax
 ity of FDR for Sparse Exponential Model\, to appear.<br/> [5] Asymptotic Mi
 nimaxity of FDR for Sparse Poisson  Model\, to appear.
SUMMARY:Detecting and Estimating Sparse Mixtures
DTSTART:20021126T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:202
SEQUENCE:0
DTEND:20021119T160000
UID:2009-11-17T20:55:58-08:00_346765546@limen.stat.ucla.edu
DESCRIPTION:More than a quarter century ago\, it was suggested that galaxie
 s such as our own Milky Way may harbor massive central black holes. Assumin
 g that gravity is the dominant force\, the motion of the stars and gas in t
 he vicinity of the putative black hole offers a robust method for establish
 ing definitive proof\, either for or against\, its existence. Since 1995\, 
 we have tracked the motion of 100 stars within several light years of the c
 enter of our Galaxy with the W. M. Keck 10-meter telescope. Although our Ga
 laxy was neither the first nor an obvious candidate for a central supermass
 ive black hole\, the results of this experiment has made the Milky Way one 
 of the strongest cases for a black hole in the million solar mass range.
SUMMARY:Unveiling a Black Hole at the Center of Our Galaxy
DTSTART:20021119T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:204
SEQUENCE:0
DTEND:20021105T160000
UID:2009-11-17T20:55:58-08:00_434490328@limen.stat.ucla.edu
DESCRIPTION:Markov Chain Monte Carlo (MCMC) techniques revolutionized stati
 stical practice in the 1990s by providing an essential toolkit for making t
 he rigor and flexibility of Bayesian analysis computationally practical. At
  the same time the increasing prevalence of massive datasets and the expans
 ion of the field of data mining has created the need to produce statistical
 ly sound methods that scale to these large problems. Except for the most tr
 ivial examples\, current MCMC methods require a complete scan of the datase
 t for each iteration eliminating their candidacy as feasible data mining te
 chniques. \n\n I will present a method for making Bayesian analysis of mass
 ive datasets computationally feasible. The algorithm simulates from a poste
 rior distribution that conditions on a smaller\, more manageable portion of
  the dataset. The remainder of the dataset may be incorporated by reweighti
 ng the initial draws using importance sampling. Computation of the importan
 ce weights requires a single scan of the remaining observations. While impo
 rtance sampling increases efficiency in data access\, it comes at the expen
 se of estimation efficiency. A simple modification\, based on the "rejuvena
 tion" step used in particle filters for dynamic system models\, sidesteps t
 he loss of efficiency with a small increase in the number of data accesses.
  \n\n On simulated and real datasets we have found reductions in excess of 
 98% in the number of data accesses without a loss of Monte Carlo precision.
SUMMARY:Bayesian Analysis of Massive Datasets via Particle Filters
DTSTART:20021105T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:205
SEQUENCE:0
DTEND:20021021T160000
UID:2009-11-17T20:55:58-08:00_918337160@limen.stat.ucla.edu
DESCRIPTION:In factor screening\, often only a few factors among a large po
 ol of potential factors are active. Under such assumption of effect sparsit
 y\, in choosing a design for factor screening\, it is important to consider
  projections of the design onto small subsets of factors. I will present so
 me recent results on hidden projection properties of certain orthogonal arr
 ays.
SUMMARY:Projection Properties of Orthogonal Arrays
DTSTART:20021021T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:206
SEQUENCE:0
DTEND:20021016T160000
UID:2009-11-17T20:55:58-08:00_706869253@limen.stat.ucla.edu
DESCRIPTION:The problem of estimating the number of unseen species in a wil
 dlife sample has been discussed by many researchers.  Here a Bayesian appro
 ach\, originally by Boender and Rinnooy Kan (1987\, Biometrika)\, is used t
 o find the posterior distribution for the number of unseen species under a 
 multinomial-Dirichlet model.  Properties of the Bayesian model are explored
 .  In addition\, the posterior distribution for the number of unseen specie
 s is used to design a future sampling plan. For example\, we find the small
 est sample size needed to obtain representatives of all of the species with
  probability 0.9.
SUMMARY:Sample Size Calculations for Finding Unseen Species
DTSTART:20021016T150000
DTSTAMP:20091117T205558
LOCATION:53-105 Center for Health Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:207
SEQUENCE:0
DTEND:20021008T160000
UID:2009-11-17T20:55:58-08:00_374842732@limen.stat.ucla.edu
DESCRIPTION:Non diabetic levels of fasting plasma glucose (FPG) have not be
 en considered particularly relevant in the prognosis for subjects with card
 iovascular disease (CVD).  Examination of the relation of FPG to 2-year mor
 tality in the Framingham Heart Study data suggests this may need serious re
 thinking.
SUMMARY:The Relation Between Glucose and Mortality
DTSTART:20021008T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:209
SEQUENCE:0
DTEND:20020528T160000
UID:2009-11-17T20:55:58-08:00_232610665@limen.stat.ucla.edu
DESCRIPTION:Kolmogorov's axiomatization of probability includes the familia
 r ratio formula for conditional probability: \n\n (RATIO) <span class='math
 '>P(A|B) = P(A \cap B)/P(B)</span> (provided <span class='math'>P(B) > 0</s
 pan>) \n\n Call this the ratio analysis of conditional probability. It has 
 become so entrenched that it is often referred to as the definition of cond
 itional probability. I argue that it is not even an adequate analysis of th
 at concept. I adduce three main classes of problem cases: \n\n * Cases in w
 hich <span class='math'>P(B) = 0</span> \n\n * Cases in which <span class='
 math'>P(A \cap B)</span> and <span class='math'>P(B)</span> are vague \n\n 
 * Cases in which <span class='math'>P(A \cap B)</span> and <span class='mat
 h'>P(B)</span> are undefined \n\n I marshal several examples of such cases 
 from scientific and philosophical practice. I conclude more positively: we 
 should reverse the traditional direction of analysis. Conditional probabili
 ty should be taken as the primitive notion\, and unconditional probability 
 should be analyzed in terms of it.
SUMMARY:What Conditional Probability Could Not Be?
DTSTART:20020528T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:210
SEQUENCE:0
DTEND:20020521T160000
UID:2009-11-17T20:55:58-08:00_57139820@limen.stat.ucla.edu
DESCRIPTION:Suppose a density function is modeled as a possibly infinite mi
 xture of densities in a given parametric family. The accuracy of k-term app
 roximations and likelihood-based estimators is examined using the Kullback-
 Leibler risk.
SUMMARY:Information-Theoretic Bounds for Mixture Density Estimation
DTSTART:20020521T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:211
SEQUENCE:0
DTEND:20020514T160000
UID:2009-11-17T20:55:58-08:00_228721444@limen.stat.ucla.edu
DESCRIPTION:This talk will give an overview of models for codon substitutio
 n and rate variation in molecular phylogeny. Particular attention will be p
 aid to a) reversibility\, b) acceptance and rejection of proposed codon cha
 nges\, c) varying rates of evolution among codon sites\, and d) the interac
 tion of these sites in determining evolutionary rates. To accommodate spati
 al variation in rates\, Markov random fields rather than Markov chains are 
 convenient. Because these innovations complicate maximum likelihood estimat
 ion in phylogeny reconstruction\, it is necessary to formulate new algorith
 ms for the evaluation of the likelihood and its derivatives with respect to
  the underlying kinetic\, acceptance\, and spatial parameters. To derive th
 e most from maximum likelihood analysis of sequence data\, it useful to com
 pute posterior probabilities assigning residues to internal nodes and evolu
 tionary rate classes to codon sites. It is also helpful to search through t
 ree space in a way that respects accepted phylogenetic relationships. Illus
 trations from the HIV genome\, the mitochrondrial genome\, and the beta glo
 blin gene will be given.
SUMMARY:Codon and Rate Variation Models in Molecular Phylogeny
DTSTART:20020514T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:212
SEQUENCE:0
DTEND:20020507T160000
UID:2009-11-17T20:55:58-08:00_201377963@limen.stat.ucla.edu
DESCRIPTION:The problem of characterizing and detecting recurrent sequence 
 patterns such as substrings or motifs and related associations or rules is 
 variously pursued in order to compress data\, unveil structure\, infer succ
 inct descriptions\, extract and classify features\, etc. In Molecular Biolo
 gy\, exceptionally frequent or rare words in bio-sequences have been implic
 ated in various facets of biological function and structure. The discovery\
 , particularly on a massive scale\, of such patterns poses interesting meth
 odological and algorithmic problems\, and often exposes scenarios in which 
 tables and synopses grow faster and bigger than the raw sequences they are 
 meant to encapsulate. In previous study\, the ability to succinctly compute
 \, store\, and display unusual substrings has been linked to a subtle inter
 play between the combinatorics of the subwords of a word and local monotoni
 cities of some scores used to measure the departure from expectation. In th
 is talk\, we carry out an extensive analysis of such monotonicities for a b
 roader variety of scores. This supports the construction of data structures
  and algorithms capable of performing global detection of unusual substring
 s in time and space linear in the subject sequences\, under various probabi
 listic models. (joint work with Alberto Apostolico [Purdue & Padova]\, Mary
  Ellen Bock [Purdue]).
SUMMARY:Monotony of Surprise and Large-Scale Quest for Unusual Words
DTSTART:20020507T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:215
SEQUENCE:0
DTEND:20020416T160000
UID:2009-11-17T20:55:58-08:00_693104510@limen.stat.ucla.edu
DESCRIPTION:We present flexible\, fully Bayesian models for the covariance 
 structure of unbalanced multivariate repeated measures data. A typical sett
 ing involves measurements on many subjects taken at any of a large number o
 f possible times. Unstructured covariance matrices have too many parameters
  to be fit to this sort of data\, so researchers have typically relied on s
 tructured covariance matrices which depend on a small set of unknown parame
 ters. We introduce prior distribution families for unstructured covariance 
 matrices that allow the data to determine a compromise between unstructured
  and parametric matrices. Applications to data from the UCLA Brain Injury R
 esearch Center are discussed.
SUMMARY:Models for the Covariance Matrix of Multivariate Longitudinal and R
 epeated Measures Data
DTSTART:20020416T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:216
SEQUENCE:0
DTEND:20020409T160000
UID:2009-11-17T20:55:58-08:00_367883297@limen.stat.ucla.edu
DESCRIPTION:Incidental mortalities of dolphin can occur during fishing oper
 ations of the international purse-seine fishery for tunas in the eastern tr
 opical Pacific Ocean. A considerable reduction in the number of marine mamm
 al mortalities has occurred since the early 1970's as a result of fleet-wid
 e implementation of fishing techniques and modifications to fishing gear wh
 ich minimize marine mammal bycatch. In the last decade\, efforts to further
  reduce dolphin mortalities have focused on individual-vessel mortality quo
 tas\, thus holding individual vessels accountable for dolphin mortalities. 
 The increased pressure on fishermen caused by individual-vessel quotas has 
 resulted in increased pressure on tuna vessel observers who are responsible
  for collecting dolphin mortality data. We present a preliminary analysis o
 f mortality data which allows us to explore the potential presence of misre
 porting\, and perhaps the possible bounds of this effect. Misreporting is i
 dentified by taking into consideration observer-captains interactions. We d
 iscuss the possibility of a more sophisticated analysis that would allow fo
 r statistically rigorous data quality control.
SUMMARY:Data Quality Control: A Fisheries Example
DTSTART:20020409T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:217
SEQUENCE:0
DTEND:20020312T160000
UID:2009-11-17T20:55:58-08:00_511234517@limen.stat.ucla.edu
DESCRIPTION:The data used in empirical social-science research\, especially
  in face to face surveys\, are often collected in multistage cluster sample
 s. The relative homogeneity of the clusters selected in this could lead to 
 design effects at the sampling stage. Interviewers tend to further homogeni
 ze answers within the sampling points. The study presented here was designe
 d to separate the two sources. Multilevel models had been used to separate 
 interviewer effects and sampling-point effects. Even though one would assum
 e that a design effect in questions of "fear of crime in the neighborhood" 
 should be due to spatial homogeneity it turned out that\, for most of the i
 tems\, the interviewer takes a far greater share of the homogenized effect 
 than the spatial clustering does.
SUMMARY:Separating Interviewer Effects and Sampling Point Effects using Int
 erpenetrated Samples with Respect to Fear of Crime Indicators
DTSTART:20020312T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:218
SEQUENCE:0
DTEND:20020226T160000
UID:2009-11-17T20:55:58-08:00_664292803@limen.stat.ucla.edu
DESCRIPTION:I will describe an approach to reading character sequences in c
 luttered environments where segmentation does not seem to be an option. A c
 omprehensive and intractable Bayesian model for the data is defined\, invol
 ving complex priors on the arrangement of the characters\, complex data mod
 els of the image given the character poses etc. A sequence of coarse to fin
 e approximations is then described each of which can be very efficiently ma
 ximized to obtain a small number of candidate solutions. At the final stage
  the prior is brought in to help sort out among a few candidate interpretat
 ions. Training involves very simple estimates of probabilities of local fea
 tures\, and is based on character templates only.
SUMMARY:Coarse to Fine Templates for Reading License Plates
DTSTART:20020226T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:219
SEQUENCE:0
DTEND:20020220T160000
UID:2009-11-17T20:55:58-08:00_412786062@limen.stat.ucla.edu
DESCRIPTION:This talk presents a procedure for detecting heterogeneity in a
  sample with respect to a given model. It can be applied to find if a univa
 riate sample or a multivariate sample has been generated by different distr
 ibutions\, or if a regression equation is really a mixture of different reg
 ression lines. Based on some special features of cross-validating predictiv
 e distributions\, the idea of the procedure is first to split the sample in
 to more homogeneous groups and then recombine the observations in order to 
 form homogeneous clusters. The proposed procedure can be applied to find he
 terogeneity in any statistical model. The performance of the procedure is i
 llustrated in univariate\, multivariate and linear regression problems.
SUMMARY:The SAR Procedure: A Diagnostic Analysis of Heterogeneous Data
DTSTART:20020220T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:225
SEQUENCE:0
DTEND:20020122T160000
UID:2009-11-17T20:55:58-08:00_205421816@limen.stat.ucla.edu
DESCRIPTION:A Bayesian approach to the classification problem is proposed i
 n which random partitions play a central role. It is argued that the partit
 ioning approach has the capacity to take advantage of a variety of large-sc
 ale spatial structures\, if they are present in the unknown regression func
 tion <span class='math'>f_0</span>. An idealized one-dimensional problem is
  considered in detail. The proposed nonparametric prior is found to provide
  a consistent estimate of the regression function in the <span class='math'
 >L^p</span> topology\, for any <span class='math'>1 \leq p < \infty </span>
 \, and for arbitrary measurable <span class='math'>f_0:[0\,1] \rightarrow [
 0\,1]</span>. An MCMC implementation is outlined and simulation experiments
  are conducted to show that the proposed estimate compares favorably with C
 ART and bagged CART estimates. A generalized prior is discussed which emplo
 ys a random Voronoi partition of the covariate-space. The resulting estimat
 e displays promise on a two-dimensional problem\, and extends with a minimu
 m of additional computational effort to arbitrary metric spaces.
SUMMARY:Bayesian Nonparametric Classification
DTSTART:20020122T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:223
SEQUENCE:0
DTEND:20020402T160000
UID:2009-11-17T20:55:58-08:00_7636096@limen.stat.ucla.edu
DESCRIPTION:Random field regression models are a popular choice for modelin
 g data from experiments with high signal-to-noise ratios\, for example comp
 uter experiments. The idea is to model the output Y from an experiment with
  factors <span class='math'>X_1\,\ldots\,X_k</span> as the realization of a
  Gaussian process with covariance function <span class='math'>C(X_1\,X_2)</
 span>. Typically the covariance function will depend on a number of unknonw
 n parameters that must be estimated from the data. Responses at further inp
 ut settings can be estimated by their BLUP's. These models have proven succ
 essful in applications. However\, they can be difficult to interpret. In th
 is talk we show that random field regression models have a natural interpre
 tation in terms of Bayesian regression models. We present some simple data 
 analytic tools that make it possible to "discover" the associated regressio
 n model.
SUMMARY:Data Analytic Methods for Understanding Random Field Regression
DTSTART:20020402T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:221
SEQUENCE:0
DTEND:20020212T160000
UID:2009-11-17T20:55:58-08:00_505518722@limen.stat.ucla.edu
DESCRIPTION:If we are interested in the estimation of an unknown parameter 
 theta (this doesn't mean that we are in a parametric context) and theta doe
 s not belong to a Euclidean space. This can happen if the parameter is a de
 cision tree\, a permutation\, a phylogenetic tree etc.... How can we go abo
 ut our usual statistical procedures? I will take the example of phylogeneti
 c trees and permutation data\, both important in computational biology to s
 ee how to extend statistical procedures such as averaging\, constructing co
 nfidence regions\, bootstrapping to the more general frameworks where the g
 eometry of the space of parameters is not Euclidean. The examples presented
  come from biological situations where DNA data is used to construct phylog
 enetic trees that are the parameters for which we would like to make confid
 ence regions.
SUMMARY:Statistics in Tree Space
DTSTART:20020212T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:224
SEQUENCE:0
DTEND:20020129T160000
UID:2009-11-17T20:55:58-08:00_917209905@limen.stat.ucla.edu
DESCRIPTION:When testing multiple hypotheses it is important to assess the 
 number of false positives in some fashion. In order to accomplish this task
  we investigate the False Discovery Rate (FDR)\, and we introduce a new qua
 ntity called the positive False Discovery Rate (pFDR). We show that the pFD
 R can be written in a very simple form and has a Bayesian interpretation. W
 e also suggest a more direct approach to multiple hypothesis testing than w
 hat has traditionally been taken. Instead of fixing the error rate and esti
 mating the corresponding rejection region\, we take the opposite approach: 
 we fix the rejection region and estimate its corresponding error rate. We s
 how how this approach can be applied to the pFDR and the FDR\, resulting in
  substantial improvements to power\, interprebility\, and applicability. Th
 is methodology works particularly well for large numbers of hypothesis test
 s and is also immune to certain forms of dependence&mdash\;both of which ma
 ke it particularly applicable to DNA microarrays\, where it is often the ca
 se that thousands of dependent hypotheses are simultaneously tested.
SUMMARY:A Direct Approach to False Discovery Rates
DTSTART:20020129T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:226
SEQUENCE:0
DTEND:20020115T160000
UID:2009-11-17T20:55:58-08:00_790763856@limen.stat.ucla.edu
DESCRIPTION:In a classic two-sample problem we might use Wilcoxon's statist
 ic to test for a difference between Treatment and Control subjects. The ana
 logous microarray experiment yields thousands of Wilcoxon statistics\, one 
 for each gene on the array\, and we would be faced with a difficult simulta
 neous inference situation. We will discuss two inferential approaches to th
 is problem: an empirical Bayes method that requires very little a priori Ba
 yesian modeling\, and the frequentist method of ""False Discovery Rates"" p
 roposed by Benjamini and Hochberg in 1995. It turns out that the two method
 s are very closely related and can be used together to produce sensible sim
 ultaneous inferences.
SUMMARY:Microarrays\, Empirical Bayes Methods\, and False Discovery Rates
DTSTART:20020115T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:229
SEQUENCE:0
DTEND:20011204T160000
UID:2009-11-17T20:55:58-08:00_730752728@limen.stat.ucla.edu
DESCRIPTION:Textures are studied extensively in computer vision and graphic
 s\, and statistics plays an important role in this area. Some background on
  the main problems and methods of texture analysis will be given. Two examp
 les will be discussed in detail to show how a statistical framework can hel
 p understand\, justify and improve existing heuristic algorithms. The first
  example is a texture synthesis algorithm which produced very good visual r
 esults with a simple heuristic resampling scheme. We formalize the scheme a
 nd show that it leads to consistent distribution estimates under mixing con
 ditions. The second example deals with texture classification based on a he
 uristic distance between distributions called Earth Mover's distance. We sh
 ow that this distance is equivalent to the Mallows metric\, and demonstrate
  its connection to an approximate Bayes rule for this classification proble
 m. The connection suggests a modification to the metric that significantly 
 improves classification results.
SUMMARY:Texture from a Statistical Perspective
DTSTART:20011204T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:231
SEQUENCE:0
DTEND:20011120T160000
UID:2009-11-17T20:55:58-08:00_86662439@limen.stat.ucla.edu
DESCRIPTION:In much psychological questionnaires one is confronted with the
  phenomenon of missing item responses. Item nonresponse can cause many prob
 lems in consecutive multivariate analyses of data. This presentation discus
 ses person mean imputation when factor analysis of the data is envisaged. F
 irst\, some practical considerations are presented. New results for bias of
  imputed scores are presented. Simulation results suggest that EM methods a
 nd methods using the mean per person work well. Next\, we turn to mathemati
 cal implications of person mean imputation in the one-factor model. We requ
 ire that imputation of the person mean does not change the structure of the
  covariance matrix. To do so\, we first have to make assumptions regarding 
 the nonresponse pattern and mechanism. Some theorems are presented which fo
 llow from the covariance structure maintaining restriction given the assump
 tions. Once it is clear what properties the dispersion matrix has\, factor 
 analysis is performed on the covariance matrix. A theorem is presented whic
 h states that\, under the restriction of covariance structure maintaining p
 roperty for person mean imputation\, factor analysis results in as many neg
 ative unique variances as there are imputed variables.
SUMMARY:Practice and Theory for Imputation in Factor Analysis of Questionna
 ire Data with Item Nonresponse
DTSTART:20011120T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:227
SEQUENCE:0
DTEND:20020108T160000
UID:2009-11-17T20:55:58-08:00_135699853@limen.stat.ucla.edu
DESCRIPTION:As part of its long-standing interest in breaking down the boun
 daries separating disciplinary genres\, the Brooklyn Academy of Music devel
 oped a year-long artist-in-residency program in conjunction with Lucent Tec
 hnologies. The Arts in Multimedia (AIM) project paired artists working in n
 ew media and communications technology with scientists at Bell Laboratories
 . In 2000\, sound artist Ben Rubin (EAR Studio\, New York City) and I recei
 ved an AIM award for our proposal entitled "Ear to the Ground" (ETG). As a 
 collaboration between a statistician and an artist\, we are looking to crea
 te sound representations that further the arts of data analysis\, discovery
 \, and expression. We plan to establish a series of "listening posts"\, poi
 nts in the physical world or on the Web\, where people can engage large\, o
 ften abstract data streams\, bringing an awareness and understanding of the
 se data to the general public. By creating these installations\, we also ho
 pe to establish general principles for the public display of complex data. 
 \n\n ETG's first listening post provided a sonification of the traffic on a
  Web site. The richness of this representation is controlled by the number 
 of visitors accessing "detailed" content (Web pages or other documents). Th
 e presence of each active visitor contributes to the loudness and tonal bal
 ance of a low-register drone. Each major section of the Web site is assigne
 d a different pitch\, the relative volume of which is increased as more vis
 itors browse content from that area. Requests for more informative content 
 deep in the site are represented as higher-pitched pulsing tones: the faste
 r the pulses\, the more people are accessing that area\, and the higher the
  pitch\, the more detailed the content. See Hansen and Rubin (2000) and Han
 sen and Rubin (2001) for more information. \n\n ETG's second project titled
  "Listening Post" focuses on answering the question "What does the collecti
 ve voice of the Internet sound like?" Countless others are with you when yo
 u browse the Web\, some reading the same words at the same time\, and yet y
 ou have no way of sensing their presence. Listening Post gives voice to thi
 s vast\, silent world\, transforming collective online activity and communi
 cation into a multi-layered sound installation. At the center of this uniqu
 ely designed space is a large array of vacuum fluorescent displays (VFDs) t
 hat flash samples from tens of thousands of online exchanges in real time\,
  revealing the patterns and rhythms of people communicating with each other
 . The technical challenges implied here are considerable\; from scalable mo
 nitoring agents that continually cull new content on the Web and update var
 ious measures of activity\, to statistical natural language processing and 
 dynamic clustering schemes that allow us to track topics and extract repres
 entative phrases. \n\n Listening Post was installed as part of the Brooklyn
  Academy of Music's 2001 Next Wave Festival and was open to the public from
  December 6 -- 20\, 2001. Additional funding for this piece has been suppli
 ed by grants from the Rockefeller Foundation and the New York State Council
  for the Arts. More information about ETG can be found in a recent NY Times
  article or at the ETG Web site .
SUMMARY:Experiencing Information Systems Through Sound
DTSTART:20020108T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:233
SEQUENCE:0
DTEND:20011106T160000
UID:2009-11-17T20:55:58-08:00_63890524@limen.stat.ucla.edu
DESCRIPTION:I will discuss two difficult and important problems that arise 
 in genomics. Both problems are addressed by incorporating specific biologic
 al information into a statistical model that accounts for uncertainty regar
 ding the biological laws and uncertainty arising from measurement and sampl
 ing variation. The resulting models are complex\, and various simulation to
 ols are used to fit the models to data. Brief descriptions of the two probl
 ems follow. \n\n 1. Separation of Expression Signals from Mixed Cell Popula
 tions. When measuring gene expression in cells taken from tissue\, it is ne
 arly certain that several cell types will be present in addition to the cel
 l type of interest. I will discuss a procedure for estimating the expressio
 n signals corresponding to the biologically-pure sub-populations of cells t
 hat comprise the sample. \n\n 2. Identification of Global Regulators of Gen
 e Expression. The loosely-defined term "global regulator" refers to a relat
 ively small number of genes whose products have a wide-ranging influence on
  the state of the cell. One mechanism of action of these regulators is that
  their products bind the DNA slightly upstream of the coding region of the 
 gene whose expression they influence. Thus there is information in both gen
 e expression and genome sequence measurements regarding the identities of t
 he global regulators. I will discuss a graph-structured probability model f
 or identifying global regulators.
SUMMARY:Two Problems in Genomics that can be Addressed by Statistical Model
 ing and Simulation
DTSTART:20011106T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:235
SEQUENCE:0
DTEND:20011023T160000
UID:2009-11-17T20:55:58-08:00_803400913@limen.stat.ucla.edu
DESCRIPTION:An important problem addressed using cDNA microarray data is th
 e improvement of classification of samples into known groups (supervised cl
 assification). The solution involves several steps\, starting with data adj
 ustment\, through variable selection to development of a classification rul
 e. Unfortunately\, the strong interaction between the steps is often ignore
 d. This talk will consider supervised classification from a unified point o
 f view.
SUMMARY:Variable Selection and Pattern Recognition with Microarray Data
DTSTART:20011023T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:236
SEQUENCE:0
DTEND:20011016T160000
UID:2009-11-17T20:55:58-08:00_298615790@limen.stat.ucla.edu
DESCRIPTION:Generalized Linear Mixed Effects Models (GLMM) provide useful t
 ools for correlated and/or overdispersed non-Gaussian data. In this talk we
  consider Generalized Non-parametric Mixed Effects Models (GNMM) which rela
 x the rigid linear assumption on the conditional predictor in a GLMM. We us
 e smoothing splines to model fixed effects. The random effects are general 
 and may also contain stochastic processes corresponding to smoothing spline
 s. We show how to construct smoothing spline ANOVA decompositions for the c
 onditional predictor function. We develop an estimation procedure that uses
  penalized likelihood\, Newton-Raphson method\, stochastic approximation an
 d Markov Chain Monte-Carlo. We apply the method to a real data set.
SUMMARY:Generalized Non-parametric Mixed Effects Model
DTSTART:20011016T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:237
SEQUENCE:0
DTEND:20011009T160000
UID:2009-11-17T20:55:58-08:00_297647552@limen.stat.ucla.edu
DESCRIPTION:Classical uniform random number generators have some major defe
 cts\, such as\, short period length and lack of higher dimension uniformity
 . Deng and Lin (2000) proposed a special class of generators which is as ef
 ficient the classical generators while enjoy the property of a much longer 
 period and of a higher dimension uniformity. In this talk\, we will first r
 eview the random number generators proposed in Deng and Lin (2000)\, we the
 n present our recent results to find a class of uniform random number gener
 ators with extremely long cycle and with a property of a much higher dimens
 ion uniformity.
SUMMARY:Recent Advances on Uniform Variate Generators
DTSTART:20011009T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:238
SEQUENCE:0
DTEND:20011002T160000
UID:2009-11-17T20:55:58-08:00_437714838@limen.stat.ucla.edu
DESCRIPTION:The incredibly complex (yet robust\, efficient and elegant) fun
 ctional\, anatomical and bio-physiological organization of the brain provid
 es a rich source for developing interesting mathematical and computational 
 models. Following an introduction to the goals of brain mapping research an
 d the variety of brain-data acquisition methods we will describe a number o
 f problems and obstacles researchers in this field encounter. Among the mos
 t needed algorithms and data filters are models for: \n\n * Stereotactic da
 ta registration (alignment)<br/> * Cortical surface modeling<br/> * Tissue 
 segmentation<br/> * Skull stripping and feature extraction<br/> * Construct
 ion of population specific brain atlases<br/> * Measures of temporal/develo
 pmental changes and variability<br/> * Assessment of structural or function
 al differences \n\n Most of these problems are amenable by various statisti
 cal analysis techniques. We will talk about the efficiency of certain mini-
 max estimators used in quantitative evaluation of different image registrat
 ions and about a statistical technique for analyzing brain functional data.
SUMMARY:Modeling\, Analyzing and Interpreting Brain Imaging Data
DTSTART:20011002T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:239
SEQUENCE:0
DTEND:20010608T160000
UID:2009-11-17T20:55:58-08:00_487358211@limen.stat.ucla.edu
DESCRIPTION:After introducing a few elementary models for investment which 
 may be new\, we soon focus our interest on the following question: How shou
 ld we invest capital into a sequence of investment opportunities\, if we wa
 nt to invest in the very last\, respectively\, very best opportunity? We in
 troduce several models to study such questions. Viewing high-risk situation
 s we assume that an investment on the very best opportunity yields a lucrat
 ive\, possibly time-dependent\, rate of return\, that uninvested capital ke
 eps its risk-free value\, whereas "wrong" investments lose their value. Sev
 eral models are presented\, mainly for the so-called rank-based case. Optim
 al strategies and values are found\, also for different utility functions. 
 In particular\, we are interested in tractable models for an unknown number
  of opportunities. The major part of this talk is based on joint work with 
 Thomas S. Ferguson.
SUMMARY:Last-Chance and High-risk Investment Models
DTSTART:20010608T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:241
SEQUENCE:0
DTEND:20010529T160000
UID:2009-11-17T20:55:58-08:00_851149981@limen.stat.ucla.edu
DESCRIPTION:The images that are cast upon our retinae are not random\, but 
 rather exhibit a certain characteristic structure due to properties of the 
 natural (or even man-made) environment. In recent years\, attempts have bee
 n made to model this structure using statistical methods\, with the goal of
  relating the structure of natural scenes to the response properties of neu
 rons in the visual system. In this talk\, I will show how the receptive fie
 ld properties of neurons in the mammalian visual cortex may be accounted fo
 r in terms of an efficient coding strategy adapted to natural images. I wil
 l also discuss a general framework for exploring the response properties of
  sensory neurons based on building probabilistic models of natural scenes.
SUMMARY:Statistics of Natural Scenes and Efficient Coding
DTSTART:20010529T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:242
SEQUENCE:0
DTEND:20010515T160000
UID:2009-11-17T20:55:58-08:00_301586211@limen.stat.ucla.edu
DESCRIPTION:A single causal factor often contributes towards producing an e
 ffect\, yet is insufficient to produce it on its own. From a set of observa
 tions\, how should one determine whether or not multiple factors interact t
 o produce or prevent an effect\, rather than influence it independently? Th
 is talk will present an argument for the necessity of the explicit represen
 tation of the possible existence of unobservable causal relations in the as
 sessment procedure\, and an example of a model for evaluating the strength 
 of conjunctive causal relations that incorporates such representation. The 
 analysis will be limited to a simple kind of conjunctive causal relations\,
  those involving (1) two candidate causes and an effect\, all of which are 
 representable by binary variables\, (2) discrete effects that occur in a se
 t of distinct entities (e.g.\, college admission being granted or not to ea
 ch student in a group)\, and (3) candidate causes for which the two values 
 (the presence and absence of a factor) respectively indicate a potentially 
 causal state and a non-causal state (e.g.\, the presence of asbestos in the
  air potentially causing lung cancer\, but the absence of asbestos not caus
 ing lung cancer\, or anything else).
SUMMARY:Assessing Interactive Causal Influence
DTSTART:20010515T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:244
SEQUENCE:0
DTEND:20010501T160000
UID:2009-11-17T20:55:58-08:00_346957650@limen.stat.ucla.edu
DESCRIPTION:Stratospheric ozone plays a vital role in restricting the ultra
 violet radiation that reaches the surface of the earth\, as well as in cont
 rolling the atmospheric temperature distribution. The study of seasonal\, c
 hemical and dynamical sources of variation in ozone levels is thus an impor
 tant component of climate research. Observations from ozonesondes (balloon-
 based recording instruments) of the ozone vertical profile over a geographi
 c location may be considered as samples from vertical profile "curves" whic
 h evolve in time\, leading us to use functional data analysis methodology f
 or studying sources of variability in the ozone vertical profile. We consid
 er a class of varying-coefficient functional data models for ozone vertical
  profiles\, where the coefficients of a basis function expansion depend on 
 covariates\, such as the quasi-biennial oscillation and season. The basis f
 unction expansion results in dimension reduction with positive implications
  for computational feasibility.
SUMMARY:Stratospheric Ozone Variability: A Functional Data Analysis Study
DTSTART:20010501T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:246
SEQUENCE:0
DTEND:20010422T160000
UID:2009-11-17T20:55:58-08:00_584867959@limen.stat.ucla.edu
DESCRIPTION:Empirical Bayes regression procedures are commonly used in educ
 ational and psychological testing as extensions to latent variable models. 
 The National Assessment of Educational Progress (NAEP) is an important nati
 onal survey in the United States using such procedures. Student responses t
 o questions (items) across various subject matters (e.g.\, reading\, scienc
 e\, and music) are collected and analyzed by correlating with their backgro
 und information such as ethnicity and parental education. NAEP applies empi
 rical Bayes methods to models from item response theory. In the process\, N
 AEP uses a two-stage procedure: first\, item parameters are estimated\, the
 n an empirical Bayes methods is applied to estimate subgroup student profic
 iencies. In the second stage\, item parameters are treated as known. We fou
 nd that the effect of ignoring uncertainty in the item parameters on report
 ed NAEP outcome can be substantial. Furthermore\, the item response theory 
 model NAEP uses ignores the effect of item clustering created by the design
  of a test form. Using Markov Chain Monte Carlo method\, we simultaneously 
 estimate all parameter of an expanded hierarchical model. The unified appro
 ach allows us to assess both the clustering and the empirical Bayes effects
 .
SUMMARY:Investigating Effects Due to Item Clustering and an Empirical Bayes
  Procedure in a National Survey
DTSTART:20010422T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:247
SEQUENCE:0
DTEND:20010420T160000
UID:2009-11-17T20:55:58-08:00_97370430@limen.stat.ucla.edu
DESCRIPTION:In the eighteenth century in France\, citizens and royalty face
 d a multitude of risks\, from sexually transmitted disease to decapitation.
  An unusual data source on the French Lottery provides a window on how fina
 ncial risk was addressed in that tumultuous time\, and how the emerging cal
 culus of probabilities affected its perception. The story involves an unusu
 ally diverse cast of characters\, including Casanova and Bonaparte\, as wel
 l as some modern probability and statistical technique.
SUMMARY:Risk and the Eighteenth Century French Lottery: Napoleon Meets his 
 Chi-Square
DTSTART:20010420T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:252
SEQUENCE:0
DTEND:20010223T160000
UID:2009-11-17T20:55:58-08:00_645844095@limen.stat.ucla.edu
DESCRIPTION:In the first part of the talk\, we develop a general framework 
 for spatio-temporal modeling. At each time period\, we write the spatial me
 an function as a locally-weighted mixture of linear regressions. To incorpo
 rate temporal variation\, we allow the regression coefficients to change ov
 er time. The model is cast in a Gaussian state-space framework\, which allo
 ws us to incorporate nonstationary components such as temporal trends and s
 easonality\, and permits efficient implemention and full probabilistic infe
 rence for the parameters\, interpolations and forecasts. To illustrate the 
 methodology\, we analyze a large dataset of Venezuelan rainfall levels. In 
 the second half of the talk\, we consider the problem of ozone monitoring i
 n Mexico City. The data consist of hourly observations of ozone\, humidity\
 , NOx\, and wind velocity from a network of 19 stations. The ozone exhibits
  strong diurnal patterns and space-time interactions\, due to photochemical
  and transport processes. We develop a seasonal state-space model that inco
 rporates wind flows\, NOx and other predictor variables\, and implement it 
 using empirical Bayes methods.
SUMMARY:Dynamic Models for Spatio-Temporal Data
DTSTART:20010223T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:249
SEQUENCE:0
DTEND:20010313T160000
UID:2009-11-17T20:55:58-08:00_486454205@limen.stat.ucla.edu
DESCRIPTION:Microarray technology enables the massive measurement of mRNA a
 t the full genome scale. It opens a wide window for scientists to visualize
  the extremely complex\, yet well-orchestrated cellular activities as encod
 ed by thousands of genes in the cell. \n\n By studying the gene expression 
 data aggregated from various biological experiments involving different cel
 l species or under a variety of environmental conditions\, interesting hypo
 theses about protein functionality can be speculated. Genes with high degre
 e of expression similarity may be co-regulated by common upstream regulator
 y factors. They are likely to be functionally related and may participate i
 n common pathways. Pearson's correlation coefficient\, a simple way of desc
 ribing the strength of linear association between a pair of random variable
 s\, has become the most popular measure of gene expression similarity. It i
 s thought to conform well to the intuitive biological notion of "coexpresse
 d" genes. \n\n However\, this statistical notion of association is not soli
 d enough to describe various kinds of in vivo relationship. Indeed\, there 
 is a liquid aspect linked to it because the abundance of different mRNA spe
 cies is subject to the influence from activities of other cellular molecule
 s. To explore the liquid aspect of gene expression association\, we introdu
 ce a notion called "liquid association" (LA). This notion intends to quanti
 fy the flow in the association between a pair of genes as the expression le
 vel of a third gene varies. A pair of genes\, A and B\, is called a LAP (li
 quid associated pair) of gene C if the corresponding LA measure (in the abs
 olute value) is large. Genes which exert greater influence on others\, as j
 udged by comparing the distribution of their LA's\, are called expression m
 asters. \n\n The notion of LAP is applied to the Stanford Yeast Cell-Cycle 
 data. Top 100 positive and 100 negative expression masters are obtained fro
 m nearly 6000 Yeast ORF's. A number of these genes have been well-documente
 d in standard textbooks for their essential roles in cellular processes. Am
 ong them\, 6 are in the energy category of glycolysis and metabolism of ene
 rgy reserves (glycogen\, trehalose): PFK1 and PFK2 (coding 6-phosphofructok
 inase\, the major flux-controlling enzyme of glycolysis)\; and TPS2\, TPS1\
 , GSY1\, GLC3 (appearing\, in KEGG's chart for the starch and sucrose metab
 olism pathway\, neck to neck with alpha-alpha Trehalose-6P\, UDP-glucose\, 
 and glycogen being the intermediate in sequence). The list of expression ma
 sters to be discussed also include RAS2 (GTP binding protein\, signal trans
 ducing) and CYR1(adenylate cyclase\, catalyzing the transformation of ATP i
 nto cAMP)\, PHO5 (major phosphate-regulated secreted acid phosphatase)\, CY
 C7 (cytochrome C isoform 2\, predominant isoform during anaerobic growth)\,
  and ATP1 (the F1 alpha subunit of the F0F1 ATPase complex in mitochondria)
 . \n\n The statistical theory for LAP is based on ideas similar to SIR/PHD 
 for high dimensional data analysis. Broader application of LAP for volumino
 us data analysis in general will be discussed.
SUMMARY:In the Lap of 'LAP': A Simple Notion for Elucidating Aggregated Mic
 roarray Gene Expression Data
DTSTART:20010313T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:250
SEQUENCE:0
DTEND:20010302T160000
UID:2009-11-17T20:55:58-08:00_911620436@limen.stat.ucla.edu
DESCRIPTION:We use Bayesian methodology to make inference from deterministi
 c models while accounting for uncertainty in the inputs to the model. This 
 methodology incorporates both prior knowledge of the distributions of the m
 odel inputs and any available data on the model inputs and outputs. Inferen
 ce about the outputs\, or any function of them\, requires a sample from the
  marginal posterior distribution of the inputs. When the analytical form of
  this posterior distribution is intractable and an exact sample is difficul
 t or impossible to obtain\, the sampling importance resampling (SIR) algori
 thm of Rubin (1988) can be used to obtain an approximate sample. I will pre
 sent an application to a deterministic model for predicting polychlorinated
  biphenyl (PCB) concentration in soil. The result is a distribution of conc
 entration in soil which accounts for uncertainty in model inputs. In cases 
 where the posterior distribution is concentrated close to a ridge in a high
 -dimensional space\, SIR may be an inefficient sampling method. An alternat
 ive sampling method is Markov chain Monte Carlo (MCMC)\, but designing a su
 ccessful MCMC algorithm in such a case is an extremely delicate problem and
  generic MCMC methods are unlikely to do well. In the second part of the ta
 lk\, I will present a new MCMC algorithm which uses nearest neighbors to ad
 apt the proposal to the (unknown) posterior region.
SUMMARY:Inference for Deterministic Simulation Models
DTSTART:20010302T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:253
SEQUENCE:0
DTEND:20010220T160000
UID:2009-11-17T20:55:58-08:00_226453646@limen.stat.ucla.edu
DESCRIPTION:In many applications one models a binary response Y\, observed 
 at sites on a spatial lattice\, in terms of corresponding vectors of explan
 atory variables. A full likelihood-based approach typically requires Markov
  Chain Monte Carlo\, and may be computationally infeasible for large lattic
 es. Heagerty and Lele (1998\, JASA) recently proposed analyzing such data w
 ith a hierarchical generalized linear model\, using composite likelihood to
  estimate the regression vector and covariance parameters of the underlying
  random field. This gives a "GEE2-like" (Liang\, Zeger and Qaqish\, 1992 JR
 SSB) estimating equation. We illustrate this method\, together with some of
  the practical difficulties in its implementation\, in a detailed analysis 
 of data on vegetation change in a region of natural forest in northern Isra
 el. We then discuss two computationally simpler alternitives\, both in the 
 spirit of generalized estimating equations. \n\n This work is joint with V.
  Landsman\, Y. Carmel and R. Kadmon.
SUMMARY:A Comparison of Some Estimating Equation Techniques for Analyzing S
 patially Distributed Binary Responses
DTSTART:20010220T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:254
SEQUENCE:0
DTEND:20010216T160000
UID:2009-11-17T20:55:58-08:00_977063962@limen.stat.ucla.edu
DESCRIPTION:Many time series can be modeled as the sum of three components:
  long-time trend\, seasonal effect and background noise. The trend superimp
 osed with the seasonal effect constitute the mean of the process. The issue
  of mean stationarity is usually the first step for further statistical inf
 erence. In this talk\, we present a theory of testing and estimation for a 
 monotonic trend and the identification of seasonal effects. Testing is cast
  as a generic "change-point" problem\, or probabilistic diagnostics. The ch
 ange-point problem has been one of the central issues of statistical infere
 nce for several decades. It includes\, for example\, testing for changes in
  weather patterns and disease rates. We are mainly concerned with "a poster
 iori" testing\, using spectral analysis to determine periodic components an
 d isotonic regression to estimate the trend. A distinctive feature of our a
 pproach is that these two problems can be treated simultaneously: isotonic 
 regression gives estimators for long-time trend with negligible influence f
 rom seasonal effects.
SUMMARY:Change-point Problem
DTSTART:20010216T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:255
SEQUENCE:0
DTEND:20010213T160000
UID:2009-11-17T20:55:58-08:00_931055075@limen.stat.ucla.edu
DESCRIPTION:To analyze an experiment where human skin cells are observed in
  motion\, a state space model is introduced. This is used to quantify conce
 pts of biological interest such as the tendency of a cell to move towards a
  stimulus\, and to suggest statistical methods for a data analysis. Several
  approaches are presented for estimation of parameters and their errors in 
 non-linear state space models. In particular\, maximum likelihood estimatio
 n is compared in theory and practice with a local quadratic approximation m
 ethod proposed by le Cam. This comparison motivates the introduction of a l
 ocal smoothed likelihood approximation\, which combines the asymptotic prop
 erties of the quadratic method with some of the finite sample attractions o
 f maximum likelihood estimation.
SUMMARY:Cell Motion and Non-linear State-space Models
DTSTART:20010213T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:256
SEQUENCE:0
DTEND:20010206T160000
UID:2009-11-17T20:55:58-08:00_168388656@limen.stat.ucla.edu
DESCRIPTION:Clustered (or hierarchically structured) data result from a hie
 rarchical sampling scheme which involves variables capturing individual eff
 ects and variables capturing cluster (or group) effects. In the area of soc
 ial sciences such as psychology\, education and medicine\, it is very commo
 n that variables capturing individual and cluster effects are not observabl
 e directly. Some existing methodologies (models and associated algorithms) 
 have been available for analysis of this kind of data. But it seems to us t
 hat those existing algorithms are not easy to program or converge slowly. I
 n this talk\, I will introduce a new formulation of the model for clustered
  data analysis and an associated EM algorithm for fitting the proposed mode
 l. The performance of the EM algorithm is illustrated by a practical data s
 et. Some potential applications of the model and the associated EM algorith
 m are discussed. \n\n Joint work from Jiajuan Liang and Peter M. Bentler
SUMMARY:An EM Algorithm for Clustered Data Analysis and Its Applications
DTSTART:20010206T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:257
SEQUENCE:0
DTEND:20010130T160000
UID:2009-11-17T20:55:58-08:00_868777040@limen.stat.ucla.edu
DESCRIPTION:Nonparametric regression\, such as smoothing spline\, is widely
  used in many scientific disciplines as a valuable data-analyzing method. T
 he use of a smoother requires the choice of a smoothing parameter which\, b
 y balancing fidelity and roughness\, controls how much the smoothing is don
 e. Two popular selection criteria to choose the smoothing parameter are <i>
 C<sub>p</sub></i> and generalized maximum likelihood (GML). Each\, however\
 , has its own problem. For <i>C<sub>p</sub></i> the problem is its high var
 iability\, while for GML\, the problem is its potentially big bias. In this
  talk we propose a new selection procedure: the extended exponential (EE) c
 riterion\, which combines the strength of <i>C<sub>p</sub></i> and GML\, ye
 t avoids their weakness in that the EE criterion has (a) small variability\
 , (b) small bias. In addition to these\, it also has (c) small tendency tow
 ard under and oversmoothing. All three criteria turn out to have simple geo
 metric interpretation\, which plays a pivotal role in our finite-sample\, n
 on-asymptotic theoretical analysis. The EE criterion is also shown to be mo
 re robust against non-normality. Some large sample results will be presente
 d and compared with their finite sample counterparts.
SUMMARY:Extended Exponential Criterion: A New Selection Procedure For Scatt
 erplot Smoothers
DTSTART:20010130T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:259
SEQUENCE:0
DTEND:20010116T160000
UID:2009-11-17T20:55:58-08:00_80170016@limen.stat.ucla.edu
DESCRIPTION:A "data depth" is a measure of how deep or how central a given 
 point is with respect to a multivariate distribution. It provides a new geo
 metric approach for quantifying complex characteristics of a multivariate d
 istribution\, and gives rise to a new set of parameters for the distributio
 n. These parameters\, including location\, scale\, quantiles\, skewness\, a
 nd kurtosis\, can all be visualized by simple graphs. Furthermore\, the cen
 ter outward ranking of the sample points provided by a data depth can lead 
 to a systematic nonparametric multivariate inference scheme. Aspects of app
 lications to DD-(depth vs. depth) plots\, construction of confidence region
 s\, multivariate process control\, and safety aviation analysis will be dis
 cussed. Some applications will be demonstrated using a multivariate dataset
  of aviation surveillance results from several airlines.
SUMMARY:Data Depth as a Nonparametric Multivariate Analysis and Its Applica
 tions
DTSTART:20010116T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:213
SEQUENCE:0
DTEND:20020430T160000
UID:2009-11-17T20:55:58-08:00_143079398@limen.stat.ucla.edu
DESCRIPTION:Several months ago the University Office of the President (UCOP
 ) produced a report which through regression analysis claimed to show that 
 the SAT 1 test now taken by all UC applicants had no "predictive validity" 
 for college performance once an applicant's high school grade point average
  and SAT 2 scores were taken into account. As such\, the SAT 1 could be dro
 pped. Independently\, David Freedman (from U.C Berkeley's Department of Sta
 tistics) and I both obtained a copy of the report and asked for the underly
 ing data. Soon after\, we were both authorized by UCOP and the UC Academic 
 Senate to do an analysis not just of the UCOP data\, but the data from our 
 respective campuses. A collaboration followed that also included Philip Sta
 rk of Berkeley. In this talk\, I will briefly review the substantive and po
 litical issues and then present our findings. Anyone interested in the inte
 rmingling of statistics\, public policy\, and politics\, should be entertai
 ned.
SUMMARY:The SATs and Admission to UCLA
DTSTART:20020430T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:260
SEQUENCE:0
DTEND:20001212T160000
UID:2009-11-17T20:55:58-08:00_147163229@limen.stat.ucla.edu
DESCRIPTION:Many vision problems can be formulated as Statistical Inference
 . This talk describes three aspects of this research program being undertak
 en by my group at SKERI. Firstly\, theoretical analysis of probabilistic th
 eories of contour detection. Secondly\, empirical analysis of cues for edge
  detection to determine fundamental bounds. Thirdly\, practical application
 s to the detection and reading of informational signs.
SUMMARY:Vision as Statistical Inference
DTSTART:20001212T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:261
SEQUENCE:0
DTEND:20001205T160000
UID:2009-11-17T20:55:58-08:00_627213482@limen.stat.ucla.edu
DESCRIPTION:We want to estimate the length distribution of fractures in a r
 ock surface from a geological map. We do not fully observe the fractures be
 cause part of the rock surface is covered by vegetation\, soil and water. T
 he covered region is very irregular. In fact\, it is not convex. This means
  that we might observe several fragments of a single crack. It is quite imp
 ossible to decide from the map if two nearby fragments belong to the same f
 racture. We therefore simplify by pretending that\, instead of the fracture
 s\, the observed fragments are independent. For this simpler model we find 
 the nonparametric maximum likelihood estimator of the length distribution. 
 We then apply this estimator to the original problem and show its consisten
 cy.
SUMMARY:Statistics of a Windowed Line Segment Process
DTSTART:20001205T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:263
SEQUENCE:0
DTEND:20001121T160000
UID:2009-11-17T20:55:58-08:00_800658250@limen.stat.ucla.edu
DESCRIPTION:Many classical nonparametric methods misbehave when the regress
 ion surface is not smooth\, that is when it displays some kind of spatial i
 nhomogeneities. For instance\, images of natural scenes exhibit rapid chang
 es of light-intensity edges and\, therefore\, classical methods are not wel
 l suited for their analysis. This talk examines the estimation of such disc
 ontinuous objects from both a theoretical and a practical viewpoint. On the
  one hand\, optimal decision theoretic results will be presented for some m
 odels of images with edges\; on the other hand\, simple and concrete proced
 ures which nearly achieve the best estimation bounds will be introduced. Th
 ese procedures are based on newly developed tools like "ridgelets" and thei
 r derivatives: e.g.\, "curvelets". Several numerical examples will illustra
 te the power of these new ideas.
SUMMARY:Ridgelets and Curvelets: New Tools for the Estimation of Discontinu
 ous Functions
DTSTART:20001121T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:266
SEQUENCE:0
DTEND:20001017T160000
UID:2009-11-17T20:55:58-08:00_678494164@limen.stat.ucla.edu
DESCRIPTION:In Michael Eisen's (or GAP) framework for seriation/clustering/
 visualization of cDNA microarray data matrix\, each row represents an indiv
 idual gene while each column stands for a time point or an experiment. The 
 numerical value at each intersection of row/column measures the RNA level f
 or that particular gene/time combination in some format. Three components a
 re essential for this framework to be useful: 1) coloring system to code nu
 merical numbers into color dots such that relative close numbers have simil
 ar visual effect (red/green for over/under expression)\; 2) proximity matri
 ces for the similarity (distance) among genes and experiments (Pearson corr
 elation for example)\; 3) algorithms for finding clustering patterns and se
 riation (tree/SOM). When the data entry is of categorical nature (clusterin
 g of subject/phenotype and subject/genotype problems)\, there is no simple 
 solution to the first two components: 1) color and 2) proximity. The propos
 ed method incorporates Homogeneity Analysis (Dual Scaling) into GAP (genera
 lized association plots) to find solutions for the coloring system and prox
 imity matrices. Tree clustering algorithms or the GAP seriation/clustering 
 can the be use to rearrange the categorical raw data matrix for visualizati
 on purpose.
SUMMARY:Information Visualization for Categorical Data Structure with Appli
 cations in Bioinformatics
DTSTART:20001017T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:267
SEQUENCE:0
DTEND:20001010T160000
UID:2009-11-17T20:55:58-08:00_326430914@limen.stat.ucla.edu
DESCRIPTION:Due to the inherent ill-posedness of statistical inverse proble
 ms\, the reconstructed images of positron emission tomography (PET) without
  regularization will have noise and edge artifacts. This is the limit of PE
 T\, which can not be resolved from the improvement of instrumental designs.
  In order to have better reconstructed images\, it is necessary to borrow t
 he strength from the related information from expertise or other tomography
  systems\, such as X-ray CT scan\, MRI\, and so forth. The correlated bound
 ary information may offer the useful information in reducing the noise and 
 edge artifacts. However\, the boundary information may be incomplete or inc
 orrect since the anatomy boundaries are different from the functional ones.
  Thus\, cross-reference is important to make use the boundary information w
 isely. In this talk\, we will present the cross-reference reconstruction me
 thods for the weighted least square and maximum likelihood estimates. Compu
 tational improvements by different algorithms and computer clusters will be
  also addressed. Empirical studies are performed at the PET system of Veter
 an General Hospital-Taipei. (2) Image segmentation is a fundamental and imp
 ortant step for image analysis. Tremendous efforts have been made to develo
 p robust and efficient segmentation techniques in literature. However\, seg
 mentation for texture images remains as a challenging and unresolved proble
 m due to its textural feature. While classical approaches may fail to give 
 successful segmentation for texture images\, human vision demonstrates its 
 incredible ability in localizing the boundaries among various textures. Enc
 ouraged by the human visual performance\, a new early vision model has been
  proposed in one of our previous works attempting to mimic the human visual
  perception. This talk will present new approaches for texture image segmen
 tation and their applications in ultrasound images that are collected in th
 e National Taiwan University Hospital.
SUMMARY:Some Statistical Analysis in PET (Positron Emission Tomography) and
  Ultrasound Images
DTSTART:20001010T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:268
SEQUENCE:0
DTEND:20001114T160000
UID:2009-11-17T20:55:58-08:00_589996779@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Evidence-confidence Calculations on Graph Models of Sequence Variat
 ion
DTSTART:20001114T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:271
SEQUENCE:0
DTEND:19991123T160000
UID:2009-11-17T20:55:58-08:00_801582373@limen.stat.ucla.edu
DESCRIPTION:Suppose that the finite population consists of N identifiable u
 nits. Associated with the i-th unit are\,  the study variable\, <span class
 ='math'>y_i</span>\, and a vector of auxiliary variables\, <span class='mat
 h'>x_i</span>. The values <span class='math'>x_1\, x_2\,\ldots\, x_N</span>
  are known for the entire population but <span class='math'>y_i</span> is k
 nown only if the i-th unit is selected in the sample. One of the fundamenta
 l questions is how to effectively use the complete auxiliary information at
  the estimation stage. In this presentation\, a unified framework is discus
 sed using a  proposed model-calibration technique. The proposed model-calib
 ration estimators can handle any linear or non-linear models and reduce to 
 the conventional calibration estimators of Deville and Särndal (1992) and/o
 r the pseudo-empirical likelihood estimators of Chen and Sitter (1999) unde
 r linear models. Some existing estimators using auxiliary information are r
 e-examined under this framework. The estimation of the finite population di
 stribution function\, using complete auxiliary information\, is also consid
 ered\, and estimators based on a generalized linear model are presented. Re
 sults of a limited simulation study on the performance of the proposed esti
 mators are reported.
SUMMARY:A Model-Calibration Approach To Using Complete Auxiliary Informatio
 n From Survey Data
DTSTART:19991123T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:272
SEQUENCE:0
DTEND:19991116T160000
UID:2009-11-17T20:55:58-08:00_546884351@limen.stat.ucla.edu
DESCRIPTION:A curious connection exists between the theory of optimal stopp
 ing for independent random variables\, and branching processes. In particul
 ar\, for the branching process <span class='math'>Z_n</span> with offspring
  distribution <span class='math'>Y</span>\, there exists a random variable 
 <span class='math'>X</span> such that the probability <span class='math'>P(
 Z_n=0)</span> of extinction of the branching process in generation <span cl
 ass='math'>n</span> is equal to the value obtained by optimally stopping th
 e sequence <span class='math'>X_1\,\ldots\,X_n</span>\, where these variabl
 es are i.i.d distributed as <span class='math'>X</span>. This correspondenc
 e furnishes a simple 'stopping rule' method for computing various character
 istics of branching processes\, including rates of convergence of the <span
  class='math'>n^{th}</span> generation's extinction probability to the even
 tual extinction probability\, for the supercritical\, critical and subcriti
 cal Galton-Watson process.
SUMMARY:A Curious Connection Between Branching Processes and Optimal Stoppi
 ng for Independent Random Variables
DTSTART:19991116T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:269
SEQUENCE:0
DTEND:19991214T160000
UID:2009-11-17T20:55:58-08:00_290085087@limen.stat.ucla.edu
DESCRIPTION:Kernel density estimation has been used with great success with
  data that fit the usual iid framework. The methods for iid data\, however\
 , do not directly apply to data from stratified multistage samples. We deve
 lop and present finite-sample and asymptotic properties of a modified densi
 ty estimator\; this estimator incorporates both the sampling weights and th
 e kernel weights.  We apply the estimator to data from the U.S. National Cr
 ime Victimization Survey\, and show how it can be used to explore and test 
 some current sociobiological and legal theories about rape.
SUMMARY:Nonparametric Density Estimation for Survey Samples
DTSTART:19991214T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:270
SEQUENCE:0
DTEND:19991207T160000
UID:2009-11-17T20:55:58-08:00_103420468@limen.stat.ucla.edu
DESCRIPTION:In the quickest detection problem it is desirable to detect a c
 hange in the probability distribution with as small delay as possible under
  the given rate of false alarms. It is known that the classical cusum proce
 dure of Page (1954) and the randomized version of the Shiryaev-Roberts proc
 edure proposed by Pollak (1986) are optimal for the problem of minimizing t
 he average detection delay in the class of procedures with a pre-specified 
 mean time to false alarm when observations are i.i.d. in pre-change and pos
 t-change modes. In this talk\, I will consider a change-point detection pro
 blem for general statistical models with possibly dependent and/or non-iden
 tically distributed observations. I will show that under certain general co
 nditions the two conventional change detection procedures asymptotically mi
 nimize the trade-off between any positive moment of the detection delay and
  false alarms when the rate of false alarms is low. In addition to modifica
 tions of the standard change-point detection procedures\, we consider their
  adaptive versions for the case where the "baseline" distribution is known 
 but the post-change distribution depends on a set of unknown parameters. Th
 e proposed adaptive detection procedures are also shown to be first-order a
 symptotically optimal with respect to any positive moment of the detection 
 delay in the worst-case scenario when the average run length to false alarm
  tends to infinity. The corresponding adaptive detection procedures are muc
 h simpler in implementation compared to the "mixtures" and the generalized 
 likelihood ratio procedures\, which are known to be asymptotically optimal 
 in the i.i.d. case. \n\n The results that I am going to present generalize 
 change-point detection theory far beyond simple i.i.d. models as well as su
 ggest new optimality criteria which are more appropriate for general stocha
 stic models. In addition\, the proposed detection algorithms have manageabl
 e complexity\, can be implemented on-line\, and yet\, are nearly optimal un
 der the different performance criteria reasonable for applications.
SUMMARY:Extended Asymptotic Optimality of certain Change-Point Detection Pr
 ocedures
DTSTART:19991207T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:276
SEQUENCE:0
DTEND:19991019T160000
UID:2009-11-17T20:55:58-08:00_261647419@limen.stat.ucla.edu
DESCRIPTION:I present a Bayesian approach to modeling of appearance and dis
 appearance times of species in the fossil record.  I consider a sampling sc
 heme where the fossil record is sampled at discrete time points.  The model
 s are applied to forams and trilobytes. \n\n Joint work with Charles Marsha
 ll\, Harvard University and Sanjib Basu\, Northern Illinois University.
SUMMARY:Statistical Modeling of a Fossil Record
DTSTART:19991019T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:275
SEQUENCE:0
DTEND:19991026T160000
UID:2009-11-17T20:55:58-08:00_249911747@limen.stat.ucla.edu
DESCRIPTION:Multivariate Adaptive Splines for Analysis of Longitudinal Data
  (MASAL) will be described. For illustration\, I apply MASAL for analyzing 
 growth curves. The method is a flexible and powerful tool for building grow
 th curve models as well as for exploring general longitudinal data and repe
 ated measures. In addition to their clinical importance\, the data used in 
 the analysis are statistically interesting because they represent a broad c
 lass of data that can have very complex structures.  A detailed analytic st
 rategy is presented to demonstrate all necessary steps for selecting a spli
 ne model to fit the growth curve data. In addition\, I will present how to 
 model the covariance structure when the measurements are obtained with an i
 rregular schedule.
SUMMARY:Analyzing Growth Curves using Adaptive Splines (MASAL)
DTSTART:19991026T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:278
SEQUENCE:0
DTEND:19991008T160000
UID:2009-11-17T20:55:58-08:00_619884114@limen.stat.ucla.edu
DESCRIPTION:Permutation Tests can be applied to an extremely wide variety o
 f problems\, as they require only the assumption of exchangeability under t
 he hypothesis. Permutation tests are exact and competitive with or superior
  in power to parametric tests\, providing most powerful unbiased tests cond
 itioned on the order statistics of the samples.  Though they fail to be exa
 ct in three common situations: 1) comparing variances\, 2) analyzing intera
 ctions\, and 3) testing multiple regression coefficients\, simulations sugg
 est they still may be asymptotically exact.  Can we prove this?
SUMMARY:Asymptotic Exchangeability
DTSTART:19991008T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:279
SEQUENCE:0
DTEND:19990607T160000
UID:2009-11-17T20:55:58-08:00_489607932@limen.stat.ucla.edu
DESCRIPTION:Researchers in the social and behavioral sciences sometime deal
  with complex models involving nonlinear relationships among latent variabl
 es\, such as an interaction between explanatory variables. Kenny-Judd (1984
 ) formulated the first nonlinear structural equation model. This model has 
 been used by researchers to investigate different approaches to estimate su
 ch model. Among others the Two Stage Least Squares method (TSLS) proposed b
 y Bollen (1995) and the Full Information methods by Joreskog and Yang (1996
 ) have been more widely applied to empirical data sets. In this presentatio
 n the TSLS and one of the full information methods\, namely\, Maximum Likel
 ihood method (ML)\, are reviewed and compared. Also\, the correction  of as
 ymptotic Standard error and Chi-squares of ML will be briefly discussed.
SUMMARY:Comparisons of the ML and TSLS Estimators for Interaction Model
DTSTART:19990607T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:280
SEQUENCE:0
DTEND:19990510T160000
UID:2009-11-17T20:55:58-08:00_529869692@limen.stat.ucla.edu
DESCRIPTION:Recent advances in molecular biology are making it relatively e
 asy to sample molecular variability in natural populations\, either directl
 y at the DNA level or indirectly via polymorphism detection methods like SN
 P arrays. Given these massive amounts of data\, probabilistic and statistic
 al methods have become very important for their analysis. Among the approac
 hes that are proving useful are the so-called `coalescent methods\,' which 
 model the genealogical relationships among the chromosomes in the sample. I
 nference in coalescent models has proved to be quite difficult and has led 
 to a number of computer-intensive approaches. In this talk I describe one a
 pproach to estimating the age of a Unique Event Polymorphism (UEP)\, for ex
 ample a deletion in a region of DNA or a mutation thought responsible for a
  particular disease. The method\, developed with Lada Markovtsova and Paul 
 Marjoram\, uses a Markov chain Monte Carlo algorithm that samples a class o
 f genealogical trees. A number of open problems will be discussed. \n\n [No
  particular knowledge of molecular biology is needed\, as the talk will be 
 self-contained]
SUMMARY:The Age of a Disease Gene: An MCMC Problem in Genetics
DTSTART:19990510T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:282
SEQUENCE:0
DTEND:19990419T160000
UID:2009-11-17T20:55:58-08:00_315038658@limen.stat.ucla.edu
DESCRIPTION:Six different samples were asked to value the identical environ
 mental good\, each via a different elicitation method:  an actual purchase 
 decision\, discrete choice\, open-ended format\, payment card\, and multipl
 e-bounded formats\, and a stated choice among an extended set of five alter
 natives.  Using a  common (a) underlying indirect utility function and (b) 
 stochastic structure\, we pool all of the data in one unified model. Choice
 s under the different elicitation methods are entirely compatible with the 
 same underlying set of homogeneous preferences\, providing heteroscedastic 
 errors across methods are permitted. Identical preferences (procedural inva
 riance) means identical expected willingness-to-pay.
SUMMARY:Alternative Nonmarket Value-Elicitation Methods: Are the Underlying
  Preferences the Same?
DTSTART:19990419T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:285
SEQUENCE:0
DTEND:19990517T160000
UID:2009-11-17T20:55:58-08:00_678584350@limen.stat.ucla.edu
DESCRIPTION:In generalized local linear models\, the GLM linear component i
 s replaced by a local polynomial estimator. This requires a bandwidth to be
  selected. In the asymptotically optimal bandwidth expression some terms de
 pend on the unknown regression function. The "rule of thumb" estimates thes
 e terms by fitting a global polynomial. We analize the performance and limi
 tations of this rule in a Montecarlo study covering Bernoulli-logit and Poi
 sson-Log models. To solve the problems that arise we devise a block method 
 where the number of blocks is automatically selected by an information crit
 erion.
SUMMARY:Fast Bandwidth Selection for Generalized Local Linear Models
DTSTART:19990517T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:287
SEQUENCE:0
DTEND:19990301T160000
UID:2009-11-17T20:55:58-08:00_428912404@limen.stat.ucla.edu
DESCRIPTION:Spectral analysis of stationary processes has a long history of
  development and interest in both theory and applications. We consider a cl
 ass of possibly nonstationary processes with a particular form of Fourier r
 epresentation whose spectral mass concentrated  on a number of straight lin
 es. Estimation of spectral density functions on these lines are considered.
  The asymptotic behavior of the bias and covariance of spectral estimates i
 s described. Some examples of such processes are presented as well as some 
 computational examples. The results include those obtained for stationary\,
  periodically stationary and almost periodically stationary processes as sp
 ecial cases.
SUMMARY:Line Spectrum Estimation of Harmonizable Processes
DTSTART:19990301T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:288
SEQUENCE:0
DTEND:19990226T160000
UID:2009-11-17T20:55:58-08:00_387736371@limen.stat.ucla.edu
DESCRIPTION:I'll describe how by a route starting with statistical consider
 ations we ended up with a definition of motifs in protein subfamilies which
  matched functional characterizations and plausible biological inferences a
 nd resulted in prediction of the outcome of an initial site directed mutage
 nesis experiment.
SUMMARY:Recognition of Highly Conserved Patterns in Protein Sequences: Anal
 ysis of Phycobiliproteins
DTSTART:19990226T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:289
SEQUENCE:0
DTEND:19990222T160000
UID:2009-11-17T20:55:58-08:00_887035369@limen.stat.ucla.edu
DESCRIPTION:Ecological inference&mdash\;determining behavior of individuals
  or subgroups from aggregate data&mdash\;is a statistical mainstay in litig
 ation brought under the Voting Rights Act of 1965.  I will discuss the inst
 itutional context very briefly\, and then compare three techniques for maki
 ng ecological inferences.  The old standby is ecological regression.  More 
 recently\, Gary King has announced "A Solution to the Ecological Inference 
 Problem". The "neighborhood model" was introduced only to demonstrate the p
 ower of assumptions\, but outperforms the competition in cases where truth 
 is known.  If time permits\, I will discuss recent Supreme Court decisions 
 on voting rights. The talk is based on joint work with Steve Klein (RAND)\,
  Mike Ostland (Berkeley)\, and Mike Roberts (Berkeley). \n\n References:  E
 valuation Review\, December\, 1991.<br/> Jurimetrics\, Summer\, 1991. <br/>
  Chance Magazine\, Vol. 6\, No. 3\, 1993. <br/> JASA\, December\, 1998.
SUMMARY:Ecological Regression and Voting Rights
DTSTART:19990222T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:293
SEQUENCE:0
DTEND:19981207T160000
UID:2009-11-17T20:55:58-08:00_693469721@limen.stat.ucla.edu
DESCRIPTION:Point processes have been used to characterize images which are
  spatial collections of points.  However\, remarkably little use has been m
 ade of point processes in the analysis of more general classes of images.  
 Certain types of images may be efficiently processed as simple functionals 
 of point processes.  For example\, some piecewise constant images may be we
 ll-approximated by Dirichlet tessellations of point patterns.  The image of
  <span class='math'>n</span> pixels may then be represented by merely the l
 ocations of the points in the corresponding point pattern and the image's v
 alues at those points. This is a highly efficient technique\, with informat
 ion about a large portion of the image stored in just one point.  Under cer
 tain assumptions this corresponds to <span class='math'>O(\sqrt{n})</span> 
 processing.  This framework also leads to a convenient method of modelling 
 high-dimensional piecewise constant images as functionals of random point p
 rocesses.  Extensions to techniques for analyzing other classes of images v
 ia point processes are discussed.
SUMMARY:A Picture Worth 32 Words
DTSTART:19981207T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:291
SEQUENCE:0
DTEND:19990201T160000
UID:2009-11-17T20:55:58-08:00_125231990@limen.stat.ucla.edu
DESCRIPTION:I will discuss the following multi-decision quickest detection 
 problem. There are N independent populations which are either statistically
  identical or a change occurs in one of them at unknown point in time. It i
 s necessary to detect the change in distribution as soon as possible (after
  it occurs) and indicate which population is "corrupted". The rate of false
  alarms and misidentification rate should be controlled by given levels. I 
 propose multi-hypothesis versions of the Page and Shiryaev-Roberts procedur
 es and prove that they are asymptotically optimal as the average run length
  to false alarm goes to infinity and the probabilities of misclassification
  go to zero. Specifically\, it will be shown that under certain conditions 
 the proposed detection-identification procedures asymptotically minimize th
 e trade-off between any positive moment of the detection delay and the rate
 s of false alarms and misclassification in the worst case scenario. At the 
 same time the corresponding detection-identification procedures are computa
 tionally simple. I will also show the results of application of the develop
 ed algorithms to detection of multiple targets in a heavy clutter that appe
 ar and disappear at unknown points in time. The data used in experiments ha
 ve been obtained from the real Infrared Search and Track system. The latter
  part of the talk is joint with Boris Rozovsky and Skirmantas Kligys of Mat
 h Dept\, USC.
SUMMARY:A Multi-Decision Change-Point Detection Problem
DTSTART:19990201T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:292
SEQUENCE:0
DTEND:19990111T160000
UID:2009-11-17T20:55:58-08:00_317056846@limen.stat.ucla.edu
DESCRIPTION:Many problems of practical interest can be formulated as the es
 timation of a certain function such as a regression function\, logistic or 
 other generalized regression function\, density function\, conditional dens
 ity function\, hazard function\, or conditional hazard function. When the f
 unction of interest depends on multiple variables\, structural models such 
 as a linear model\, partly linear model\, partly linear additive model\, ad
 ditive model\, or functional ANOVA model are often used to overcome the cur
 se of dimensionality. Extended linear modeling provides a unified framework
  for such strucvtural modeling in a broad range of statistical frameworks. 
 Huang has provided a general treatment of the consistency and rates of conv
 ergence of maximum likelihood estimation in the context of concave extended
  linear modeling. We are currently working on the extension of such results
  to let the approximating space used in the modeling procedure depend on a 
 vector of parameters. So far\, we have emphasized generalized regression (i
 ncluding ordinary regression) on the one hand and approximating spaces whos
 e components are suitably regular free-knot splines and their tensor produc
 ts on the other hand. The topic of this informal talk will be the backgroun
 d and current status of this research effort.
SUMMARY:Extended Semilinear Modeling
DTSTART:19990111T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:299
SEQUENCE:0
DTEND:19981102T160000
UID:2009-11-17T20:55:58-08:00_693671800@limen.stat.ucla.edu
DESCRIPTION:We analyze several earthquake catalogs which include origin tim
 e (<span class='math'>T</span>)\, hypocenter (<span class='math'>R^3</span>
 )\, and second-rank seismic moment tensor for each earthquake.  The tensor 
 is symmetric\, traceless\, with zero determinant: hence it has only four de
 grees of freedom &mdash\; one for the norm of the tensor (scalar seismic mo
 ment\, <span class='math'>M</span>) and three for the orientation\, <span c
 lass='math'>SO(3)</span> (they define the focal mechanism of an earthquake)
 . An earthquake occurrence is considered as a stochastic\, tensor-valued\, 
 multidimensional\, point process: <span class='math'>M \times T \times R^3 
 \times SO(3)</span>. \n\n An earthquake occurrence exhibits scale-invariant
  properties (Kagan\, <i>Physica D</i>\, 1994): \n\n (1) earthquake size dis
 tribution is a power-law (the Gutenberg-Richter relation for magnitudes or 
 the Pareto distribution for seismic moment). Conservation of energy require
 s that the distribution should be limited on the high side\; thus we use th
 e generalized gamma distribution: <span class='math'>\phi (M) \propto M^{-1
 -\beta} \exp (-M/M_{max})</span>.<br/> Both <span class='math'>\beta</span>
  and <span class='math'>M_{max}</span> have universal values for shallow ea
 rthquakes (depth interval 0-70 km)\, occurring in continents and continenta
 l boundaries (Kagan\, <i>JGR</i>\, 1997). \n\n (2) Temporal fractal pattern
 : power-law decay of the rate of the aftershock and foreshock occurrence (O
 mori's law)\, <span class='math'>\phi (t) \propto t^{-1-\theta}</span>\; <s
 pan class='math'>\theta = 0.5</span> for shallow earthquakes. \n\n (3) Spat
 ial distribution of earthquakes is fractal: as the time span of a catalog i
 ncreases\, the correlation dimension of earthquake hypocenters (<span class
 ='math'>\delta</span>) asymptotically reaches the value 2.2 for shallow ear
 thquakes. \n\n (4) 3-D disorientation of earthquake focal mechanisms is app
 roximated by the rotational Cauchy distribution. \n\n (5) We investigate th
 e statistical properties of incremental static stress caused by earthquakes
 . Theoretical calculations\, simulations and measurements of the rotation o
 f earthquake focal mechanisms suggest that the stress in the earthquake foc
 al zones follows the Cauchy distribution which is one of the stable probabi
 lity distributions. \n\n We offer a model of random defect interaction in a
  critical stress environment which seems to explain most of the available e
 mpirical results. In the time domain\, Omori's law of foreshock/aftershock 
 occurrence and\, in general\, the temporal clustering of earthquake events\
 , is a consequence of a Brownian motion-like behavior of random stress due 
 to defect dynamics. Similarly\, presence\, evolution\, and self-organized a
 ggregation of defects in the rock medium are responsible for fractal spatia
 l patterns of earthquake faults. The Cauchy and other symmetric stable dist
 ributions govern the stress caused by these defects\, as well as the random
  rotation of focal mechanisms. The stable distributions have a power-law ta
 il (i.e.\, they are fractal and should yield fractal fault patterns). \n\n 
 We discuss several definitions and possible classifications of earthquake p
 rediction methods (Kagan\, <i>GJI</i>\, 1997). An empirical search for eart
 hquake precursors which forecast the size of an impending earthquake\, has 
 been fruitless. The most probable consequence of earthquake self-similarity
  is lack of earthquake predictability as a forecast of a specific individua
 l earthquake. Many small earthquakes occur throughout any seismic zone\, de
 monstrating that the critical conditions for earthquake nucleation are sati
 sfied almost everywhere\; apparently\, any small shock can grow into a larg
 e event. Thus\, it is likely that an earthquake has no preparatory stage.<b
 r/> Although earthquake prediction\, as popularly defined\, may well be imp
 ossible\, the seismic moment conservation principle\, combined with geodeti
 c deformation data\, offers a new way to evaluate the seismic hazard\, not 
 only for tectonic plate boundaries\, but for areas of low seismicity\, i.e.
 \, the interior of continents. Earthquake clustering with the power-law tem
 poral decay can be used to estimate the time-dependent rate of future earth
 quake occurrence.
SUMMARY:Statistical Analysis\, Modelling\, and Prediction of Earthquakes
DTSTART:19981102T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:297
SEQUENCE:0
DTEND:19981116T160000
UID:2009-11-17T20:55:58-08:00_276964995@limen.stat.ucla.edu
DESCRIPTION:An important criterion for understanding earthquakes is our abi
 lity to estimate earthquake potential: the rate of future earthquakes as a 
 function of location\, time\, magnitude\, and perhaps other variables. Pred
 icting individual earthquakes with high probability seems impossible for th
 e foreseeable future\, but probabilistic "forecasting" may be possible.  Tw
 o competing views claim forecasting skill. In one family of "deterministic"
  models\, one assumes that earthquakes will occur on identifiable fault seg
 ments. A characteristic magnitude and average displacement can be estimated
  from the length of the segment\; and the rate of earthquakes from the slip
  rate on the fault. It is often assumed that characteristic earthquakes on 
 an individual segment will be quasiperiodic\, with a lognormal or Weibull d
 istribution of interval times. The alternate "stochastic" view is that eart
 hquakes are more likely in the time\, space\, and magnitude neighborhood of
  previous earthquakes. These two views give drastically different estimates
  for the probabilities of future earthquakes. The deterministic models impl
 y frequent moderate earthquakes but never very large ones. The stochastic m
 odels imply less frequent earthquakes\, occasionally really large. Accordin
 g to one view large earthquakes are nearly impossible just after a previous
  one\, but the other view suggests that this is the most likely time for an
 other quake. \n\n How can such major questions remain unresolved? Objective
  testing is difficult because earthquake data are noisy\; their quality var
 ies substantially with time\; tectonic environments differ substantially\, 
 introducing many free parameters\; testing hypotheses with past data involv
 es arbitrary selection that invites bias\; and testing with future earthqua
 kes is slow because earthquakes are rare and hypotheses evolve with time. N
 evertheless\, global tests may possibly break ties\, or even careers\, with
 in a few years.
SUMMARY:Earthquake Likelihood: Hypotheses and Tests
DTSTART:19981116T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:300
SEQUENCE:0
DTEND:19981026T160000
UID:2009-11-17T20:55:58-08:00_9868214@limen.stat.ucla.edu
DESCRIPTION:This talk will describe our efforts at developing a common stat
 istical framework for comparing image data obtained of human brain.  The da
 ta describe both brain anatomy and function across several subpopulations o
 f subjects.  The overall goal is to create a probabilistic representation e
 nabling population comparisons both visually and statistically.  We have co
 mpared individuals from diverse groups of people such as adults\, children 
 at different developmental stages\, aged groups and groups suffering from a
  variety of neurological diseases.
SUMMARY:Brain Mapping the Structure and Function of Humans
DTSTART:19981026T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:301
SEQUENCE:0
DTEND:19981019T160000
UID:2009-11-17T20:55:58-08:00_240102048@limen.stat.ucla.edu
DESCRIPTION:Monte Carlo Arithmetic (MCA) is an extension of standard floati
 ng-point arithmetic that exploits randomness in basic floating-point operat
 ions. MCA includes random rounding&mdash\;which forces roundoff errors to b
 e randomly distributed&mdash\;and precision bounding&mdash\;which limits th
 e number of significant digits in a given value by random perturbation. Ran
 dom rounding can be used to produce roundoff errors that are truly random a
 nd uncorrelated\, and that have zero expected bias.  Precision bounding can
  be used to vary precision dynamically\, to implement inexact values (value
 s known to only a few significant digits)\, and most importantly to detect 
 catastrophic cancellation\, which is the primary way that significant digit
 s are lost in numerical computation. \n\n Randomization has both theoretica
 l and practical benefits.  It has the effect of transforming any floating-p
 oint computation into a Monte Carlo computation\, and roundoff analysis int
 o statistical analysis. Unlike much previous work in this area\, MCA makes 
 no assumptions about the resulting roundoff error distributions\, such as t
 hat they are normal.  By running a program multiple times\, one directly me
 asures the sensitivity of particular outputs to random perturbations of par
 ticular inputs.  MCA thus gives a way to implement roundoff analysis\, usin
 g random variables for roundoff errors\, so that the roundoff distributions
  can be studied explicitly.  It encourages an empirical approach to evaluat
 ing numerical quality\, and gives a way to exploit the Monte Carlo method i
 n numerical computation. \n\n MCA also generally gives a different perspect
 ive on the study of error.  For example\, while floating-point summation is
  not associative\, Monte Carlo summation is "statistically associative" up 
 to the standard error of the sum.  A statistical approach avoids anomalies 
 of floating-point arithmetic. \n\n This work summarizes ways in which MCA h
 as promise as a tool in numerical computation.  It seems particularly promi
 sing as a way for the person on the street to estimate the number of signif
 icant digits in a floating-point value\, and to experiment with the effect 
 of changing the precision used in numerical computation.  Numerical modelin
 g is becoming a part of life for more and more people\, and few of these pe
 ople either enjoy or are skilled at formal error analysis\; MCA gives them 
 a way to estimate the quality of their numerical models.
SUMMARY:Monte Carlo Arithmetic: Exploiting Randomness in Floating-point Ari
 thmetic
DTSTART:19981019T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:303
SEQUENCE:0
DTEND:19981005T160000
UID:2009-11-17T20:55:58-08:00_53372682@limen.stat.ucla.edu
DESCRIPTION:According to one judicial standard\, judgment in favor of plain
 tiff is made only if it is MORE PROBABLE THAN NOT that the defendant's acti
 on was a NECESSARY CAUSE for the plaintiff's damage (or death). \n\n Let th
 e quantity PN(x\,y) stand for the probability that some outcome y would not
  have occured if it were not for some action x\, given that the action and 
 the outcome did in fact occur. \n\n I will discuss conditions under which P
 N(x\,y) can be estimated from statistical data in observational studies\, e
 xperimental studies\, and combination of the two. \n\n Remarkably\, combine
 d studies are shown to be more informative than either strictly experimenta
 l or or  strictly observational studies. As an added bonus\, we shall uncov
 er ways of testing whether units selected for a give experimental study are
  representative of their target population. \n\n Reference: J Pearl\, "Prob
 abilities of Causation\, three counterfactual interpretations and their ide
 ntification" Tech Report  R-261\, http://bayes.cs.ucla.edu/jp_home.html
SUMMARY:Estimating Probabilities of Causation
DTSTART:19981005T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:304
SEQUENCE:0
DTEND:19980427T160000
UID:2009-11-17T20:55:58-08:00_957443382@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Resampling Marked Point Processes
DTSTART:19980427T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:305
SEQUENCE:0
DTEND:19980420T160000
UID:2009-11-17T20:55:58-08:00_249990022@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:The Current Guidelines on Hypertension are Incorrect
DTSTART:19980420T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:306
SEQUENCE:0
DTEND:19980413T160000
UID:2009-11-17T20:55:58-08:00_317388907@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Bounding Treatments Effects from Experiments in the presence of Con
 tamination and Non-Compliance
DTSTART:19980413T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:307
SEQUENCE:0
DTEND:19980406T160000
UID:2009-11-17T20:55:58-08:00_973114623@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Dilation Inequalities applied to Medical Provider Profiling
DTSTART:19980406T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:308
SEQUENCE:0
DTEND:19980309T160000
UID:2009-11-17T20:55:58-08:00_737935850@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Efficiency of Case-Cohort Sampling in the Cox Regression Model
DTSTART:19980309T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:309
SEQUENCE:0
DTEND:19980304T160000
UID:2009-11-17T20:55:58-08:00_939044259@limen.stat.ucla.edu
DESCRIPTION:Scale and translation invariance is of interest in many areas o
 f science.  Statistics of natural images show strong evidence of such invar
 iance.  In this talk I will describe a model on the origin of this phenomen
 on.  The basic assumption of our model is that images are made up of projec
 tions of objects on the camera film.  Objects are distributed in the 3D wor
 ld by a Poisson distribution.  The size of the projection of an object as w
 ell as the color intensity distribution inside the projection changes by th
 e object's distance from the camera. The entire image is approximately the 
 arithmetic sum of these projections\, therefore approximately scale and tra
 nslation invariant. This model also leads to the mathematical problem to co
 nstruct "real" scale and translation invariant distributions.  At the end o
 f the talk\, I will briefly show how to construct scale and translation inv
 ariant distributions on the space of "generalized functions".
SUMMARY:Scale and Translation Invariance: Model and Theory
DTSTART:19980304T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:310
SEQUENCE:0
DTEND:19980303T160000
UID:2009-11-17T20:55:58-08:00_699532265@limen.stat.ucla.edu
DESCRIPTION:A result of Meyer (1971) describes a way of transforming a poin
 t process on the line into a Poisson process. This transformation is useful
  for evaluating models for one-dimensional point processes. Past efforts to
  generalize Meyer's theorem to higher dimensions are reviewed and a new res
 ult is presented. This result is sufficiently general to apply to a wide va
 riety of multi-dimensional point processes. The corresponding model assessm
 ent technique is applied to various models for earthquake occurrences using
  a catalog of 2\,402 micro-earthquakes occurring in Parkfield\, California 
 between 1988 and 1995. Models characterized by self-exciting behavior\, inc
 luding branching models and short-term exciting long-term correcting (SELC)
  models\, are shown to offer superior fit to some simple Poisson\, renewal 
 and Markov models.
SUMMARY:Assessment of Multi-dimensional Point Process Models with Applicati
 ons to Seismology
DTSTART:19980303T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:317
SEQUENCE:0
DTEND:19980202T160000
UID:2009-11-17T20:55:58-08:00_618274036@limen.stat.ucla.edu
DESCRIPTION:Latent variable modeling is a widely used statistical method in
  medical\, social\, and behavioral sciences.  Structural equation analysis 
 using latent variables is appealing and useful in applications where theore
 tical concepts cannot be measured directly\, or where a measurement problem
  is involved.  This talk addresses two statistical topics that are underdev
 eloped in the current use of latent variable analysis.  One topic which has
  been largely ignored is the development of diagnostic methods for examinin
 g data and model fit.  Here\, graphical procedures for examining the nature
  of departure from the fitted structural model are proposed.  Another topic
  of practical interest deals with possible nonlinear relationships among la
 tent variables.  The latent nonlinearity has been recognized as an importan
 t problem in social science applications.  This talk introduces procedures 
 for fitting nonlinear structural equation models.  Statistical properties o
 f the model fitting techniques and associated parameter estimators are disc
 ussed in terms of asymptotic theory and numerical study.
SUMMARY:On Nonlinear Latent Variable Analysis
DTSTART:19980202T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:312
SEQUENCE:0
DTEND:19980218T160000
UID:2009-11-17T20:55:58-08:00_6543980@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Improving the Rate of Convergence of EM in High-dimensional Finite 
 Mixtures
DTSTART:19980218T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:313
SEQUENCE:0
DTEND:19980211T130000
UID:2009-11-17T20:55:58-08:00_214803415@limen.stat.ucla.edu
DESCRIPTION:Statistical modeling and analysis have been applied to differen
 t music related fields. One of them is sound synthesis and analysis. Sound 
 can be represented as a real-valued function of time. This function can be 
 sampled at a small enough rate so that the resulting discrete version is al
 most as good as the continuous one. This permits us to study musical sounds
  as a discrete time series for which  many statistical techniques are avail
 able. \n\n Physical modeling suggests that many musical instruments' sounds
  are characterized by a harmonic and an additive noise signal. The noise is
  not something to get rid of rather it's an important part of the signal. I
 n this research we are interested in separating these two elements of the s
 ound. To do so we fit a local harmonic model  that tracks  changes in pitch
  and of the amplitude of the harmonics. Deterministic changes in the signal
 \, such as pitch change\, suggest that different window sizes should be con
 sidered. Various ways to choose appropiate window sizes are studied. Amongs
 t other things our analyses provides estimates of the harmonic signal and o
 f the noise signal. Different musical composition applications arise from t
 he estimates.
SUMMARY:Statistics and Music: Fitting a Local Harmonic Model to Sound Signa
 ls
DTSTART:19980211T120000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:311
SEQUENCE:0
DTEND:19980302T160000
UID:2009-11-17T20:55:58-08:00_353829563@limen.stat.ucla.edu
DESCRIPTION:Attempts to express confounding (or "effect bias") in terms of 
 statistical associations have converged on the following criterion: \n\n \n
 \n \n\n "The effect of X on Y  is not confounded if every variable Z that i
 s not affected by X is either: \n\n 1. unassociated with X\, or \n\n 2. una
 ssociated with the outcome Y within strata of X." \n\n The talk will demons
 trate that this  associational criterion does not ensure unbiased effect es
 timates\, nor does it follow from the requirement of unbiasedness. We will 
 then define a stronger notion of unbiasedness\, called "stable unbiasedness
 "\, relative to which a modified statistical criterion will be shown necess
 ary and sufficient. The necessary part will then yield a practical test for
  stable unbiasedness which\, remarkably\, does  not require knowledge of al
 l potential confounders in a problem.
SUMMARY:Why There is No Statistical Test for Confounding\, Why Many Think T
 here is\, and Why They are Almost Right
DTSTART:19980302T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:314
SEQUENCE:0
DTEND:19980211T170000
UID:2009-11-17T20:55:58-08:00_861751658@limen.stat.ucla.edu
DESCRIPTION:The elucidation of a protein's structure and function from its 
 sequence is one of the major challenges in molecular biology.  The sequence
  databases provide a valuable and rapidly growing resource to meet this cha
 llenge.  We describe a fully automated method\, PROBE\,  for the identifica
 tion and alignment of an entire protein family beginning with a single sequ
 ence.  The method also constructs a multiple sequence model that not only p
 lays a key role in the identification protein family members\, but also del
 ineate conserved motifs and conserved residues.  The core algorithm is base
 d on Bayesian statistics\, and combines the best features of two recently d
 escribed multiple sequence alignment methods\, the Gibbs sampler and HMMs. 
   Bayesian model selection procedures focus the alignment on those patterns
  which sequence data indicate are conserved across the protein family. Thes
 e Bayesian inferences are central to the model's ability to align subtly re
 lated sequences and extract distantly related family members from the datab
 ase.  Findings from this method are not rare.  Additional relationships wer
 e identified\, compared to BLAST\, in 56  of 100  randomly sampled query se
 quences. On average 4 times as many related sequences are identified.  Amon
 g the algorithm's findings to date are:  the discovery of two super familie
 s of zinc-dependent hydrolyses\, the delineation of a "swinging arm" domain
  in bacterial membrane fusion proteins\; and identification sequences dista
 ntly related to regulators of GTP binding proteins.  We will describe in so
 me detail a structural prediction for human glutamate decarboxylase. \n\n G
 lutamate decarboxylase (GAD\, EC 4.1.1.15) is the pyridoxal-5'-phosphate (p
 yridoxal-P) dependent enzyme that synthesizes  -aminobutyric acid (GABA)\, 
 the major inhibitory neurotransmitter in vertebrate brain.  Its structure i
 s unknown and its sequence had been previously reported to be unrelated to 
 any others.  PROBE identified six motifs shared by the GAD and a super-fami
 ly of 512 proteins including four proteins of known structure.  Five of the
  motifs correspond to the  /  elements and loops of the pyridoxal-P-binding
  cleft.  The sixth motif corresponds to a helical element of the small doma
 in that closes when the substrate binds.  Nineteen conserved residues were 
 identified and their functions were evaluated through comparison with struc
 tural data.  The strong conservation of the cofactor binding site in GAD in
 dicates that the unique and highly regulated transition between apo- and ho
 lo GAD is accomplished by modifications in this basic fold rather than thro
 ugh a novel folding pattern.
SUMMARY:Mining Protein Databases to Predict Structure and Function with App
 lication to Human Glutamate Decarboxylase
DTSTART:19980211T160000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:326
SEQUENCE:0
DTEND:20090528T170000
UID:2009-11-17T20:55:58-08:00_961928275@limen.stat.ucla.edu
DESCRIPTION:Rich internet applications are becoming more popular in the sta
 tistics community as a means to create expressive visual displays of comple
 x data sets thereby improving the accessibility and understanding of these 
 data.  Also\, they make good complements to our instructional applets and c
 ourse materials.
SUMMARY:Data Visualization
DTSTART:20090528T160000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:316
SEQUENCE:0
DTEND:19980209T160000
UID:2009-11-17T20:55:58-08:00_341859013@limen.stat.ucla.edu
DESCRIPTION:Stein's method has proven to be a powerful tool for deriving bo
 unds on distributional approximations. Most well known is Stein's method in
  the context of normal and Poisson approximations\; however\, recent effort
 s have included the chi-square distribution. \n\n Although chi-square appro
 ximations play an important role in theoretical statistics\, the rates of c
 onvergence that have been derived theoretically sometimes seem too slow com
 pared to practical experience. Typically these rates are derived using a mu
 ltivariate normal approximation and the triangle inequality. Here we will p
 resent a direct method for chi-square approximation as well as coupling tec
 hniques useful in applying this method. As the approach is based on Stein's
  method\, a bound on the rate of convergence can easily be obtained. As an 
 example we will consider Pearson's chi-square statistic\; moreover an appli
 cation to Friedman's statistics in nonparametric two-way analysis of varian
 ce will be outlined.
SUMMARY:Chi-square Approximations with Stein's Method
DTSTART:19980209T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:318
SEQUENCE:0
DTEND:19980125T160000
UID:2009-11-17T20:55:58-08:00_438833217@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:A Nonparametric Classification Model for Predicting Binary Outcomes
DTSTART:19980125T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:319
SEQUENCE:0
DTEND:19980112T160000
UID:2009-11-17T20:55:58-08:00_63800022@limen.stat.ucla.edu
DESCRIPTION:The Multi-angle Imaging SpectroRadiometer is a remote sensing d
 evice designed and built by JPL as part of the NASA's Earth Observing Syste
 m. Data collected will be used to characterize surface and atmospheric feat
 ures of the earth. In raw form\, the data are 36 measured radiances for eac
 h of 25.5 billion blocks into which the earth's surface is partitioned by t
 he instrument. The entire globe is covered once in about nine days\, and th
 e mission is slated to last six years. What will be done to convert these d
 ata into meaningful physical quantities and reduce data volume to managable
  size? Answers to these questions will be discussed\, with particular atten
 tion to statistical issues concerning lossy compression by spatial and temp
 oral aggregation.
SUMMARY:Data Analysis\, Aggregation and Remote Sensing
DTSTART:19980112T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:320
SEQUENCE:0
DTEND:19971208T160000
UID:2009-11-17T20:55:58-08:00_877024576@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Estimation of Simultaneous Equations Models with an Application to 
 the Demand for Fish
DTSTART:19971208T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:322
SEQUENCE:0
DTEND:19971204T160000
UID:2009-11-17T20:55:58-08:00_565364255@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:A Class of Models for Gene Localization
DTSTART:19971204T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:321
SEQUENCE:0
DTEND:19971201T160000
UID:2009-11-17T20:55:58-08:00_96167956@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Local Markov Trends and Macroeconomic Time Series: Evidence and Imp
 lications
DTSTART:19971201T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:323
SEQUENCE:0
DTEND:19971117T160000
UID:2009-11-17T20:55:58-08:00_663843071@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Three-way Subclassification for Multiple-class Discriminant Analysi
 s with Application to Handwritten Digit Recognition
DTSTART:19971117T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:325
SEQUENCE:0
DTEND:19971103T160000
UID:2009-11-17T20:55:58-08:00_349528138@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Evaluating The Placement of Inmates within California Prisons: A Re
 gression Discontinuity Approach
DTSTART:19971103T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:328
SEQUENCE:0
DTEND:20090423T170000
UID:2009-11-17T20:55:58-08:00_109727467@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:What Features Would an R Package for Intro Stats Possess?
DTSTART:20090423T160000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:327
SEQUENCE:0
DTEND:20090507T170000
UID:2009-11-17T20:55:58-08:00_694023492@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:How Can We Teach Computation in Intro Stats?
DTSTART:20090507T160000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:329
SEQUENCE:0
DTEND:20090416T170000
UID:2009-11-17T20:55:58-08:00_985604070@limen.stat.ucla.edu
DESCRIPTION:StatCrunch has traditionally been a Web based software package 
 for analyzing data. However\, recent updates to the statcrunch.com site hav
 e now opened the door for online social data analysis in a way that can hav
 e tremendous impacts on statistical education. This talk will cover basic u
 sage of the powerful StatCrunch statistical software package as well as met
 hods for optimizing the usage of the statcrunch.com site in your course.
SUMMARY:Using StatCrunch to Broaden the Horizons of an Introductory Statist
 ics Course
DTSTART:20090416T160000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:331
SEQUENCE:0
DTEND:20090402T170000
UID:2009-11-17T20:55:58-08:00_703918488@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Should We Use Large Data Sets in Intro Stats?
DTSTART:20090402T160000
DTSTAMP:20091117T205558
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:324
SEQUENCE:0
DTEND:19971110T160000
UID:2009-11-17T20:55:58-08:00_575620795@limen.stat.ucla.edu
DESCRIPTION:It can be argued that image processing will be a central techno
 logy in the information age. Much of the bandwidth on the internet is alrea
 dy taken up by transmission of image data. Applications to medicine is well
 -known and a mature industry. The problems are computationally intensive an
 d accurate and efficient methods are needed. Much of the research in image 
 processing involve techniques from mathematics and computer science and thu
 s is well suited to the missions of any Computational Science and Engineern
 g program. Image processing includes several subareas\, such as enhancement
 \, compression\, restoration\, segmentation\, etc. Our work so far has been
  focused on image restoration\, which refers to the process of recovering a
 n image contaminated by blurring and noise\, as well as image segmentation.
  \n\n Standard restoration methods involve computation in the frequency dom
 ain\, facilitated by efficient FFT and wavelet algorithms. Recently\, there
  has been a new movement towards a nonlinear partial differential equation 
 (PDE) based approach\, which is motivated by a more systematic approach to 
 restoring images with sharp edges\, as well as for image segmentation. The 
 image is diffused (denoised) according to a nonlinear anisotropic diffusion
  PDE\, designed to diffuse less near edges. Moreover\, the PDEs are designe
 d to possess certain desirable geometrical properties such as affine invari
 ance and causality. The total variation approach\, originally proposed by R
 udin\, Osher and Fatemi in 1992\, is a method in this family. It can be vie
 wed as a specific example of Tikhonov regularization using the total variat
 ion as a regularization functional. The first order Euler-Lagrange optimali
 ty condition leads to a nonlinear PDE with a convolution fitting term. \n\n
  From a computational standpoint\, the PDE formulations calls for new compu
 tational techniques which are different from the traditional frequency doma
 in and algebraic approaches. Among the computational difficulties are the h
 ighly nonlinear and singular nature of the PDEs that arise and the need to 
 invert ill-conditioned nonlinear differential-integral operators efficientl
 y. As yet\, the nonlinear diffusion models are considered somewhat expensiv
 e compared to traditional methods and much room for improvement exist. \n\n
  In my talk\, I will give an introduction to this field after which I will 
 highlight some of our work on the development of efficient numerical method
 s\, as well as new models for multi-spectral and color images\, blind decon
 volution\, etc. \n\n BIOGRAPHY \n\n Tony F-C. Chan received the B.S. degree
  in Engineering from the California Institute of Technology in 1973\, the M
 .S. degree in Aerospace Engineering from the California Institute of Techno
 logy in 1973\, and the Ph.D. degree in Computer Science from Stanford Unive
 rsity in 1978. \n\n He is currently chair of the mathematics department at 
 the University of California\, Los Angeles\, where he has been a professor 
 since 1986. His research interests include the design of efficient computat
 ional algorithms for large scale scientific computing (e.g. multigrid and d
 omain decomposition algorithms\, iterative methods\, Krylov subspace method
 s\, and parallel algorithms)\, VLSI circuit placement and PDE methods for i
 mage processing.
SUMMARY:Nonlinear PDE Models for Image Processing
DTSTART:19971110T150000
DTSTAMP:20091117T205558
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:335
SEQUENCE:0
DTEND:20080424T170000
UID:2009-11-17T20:55:58-08:00_582064697@limen.stat.ucla.edu
DESCRIPTION:Sophisticated statistical tools have made data analysis accessi
 ble to an increasingly wide variety of people\, from scientists to students
 \, demographers to historians.  Yet\, statistics is still generally regarde
 d as a difficult &mdash\; if not impossible &mdash\; topic to understand.  
 How can the powerful technology used for statistical analysis be harnessed 
 to support students' understanding of statistical concepts? At least two pr
 ojects in the last decade have taken on this challenge and designed educati
 onal environments that are both a tool that can carry out statistical analy
 ses and a tool box with which budding analysts can try out and compare a va
 riety of approaches to a statistical situation. \n\n I will discuss one of 
 these\, TinkerPlots\, a statistics education tool that can be used as early
  as middle school and at least through high school.  While I will describe 
 the educational model that TinkerPlots is based on and demonstrate some of 
 its features\, I will focus in particular on the ways in which the software
  acts as a learning context and share several examples of students and teac
 hers exploring statistical concepts using  the TinkerPlots tool box. \n\n B
 io: Andee Rubin\, Senior Scientist at TERC\, has done research and developm
 ent in the fields of mathematics\, technology\, and online learning for ove
 r 25 years. She has written curriculum\, developed and provided professiona
 l development\, and designed software and accompanying activities as well a
 s studying how students and teachers develop mathematical reasoning skills.
  Her research has focused on how students and teachers develop statistical 
 reasoning\, how video can be used to introduce ideas of movement over time 
 to middle school students\, and how mathematics education can be integrated
  into informal settings such as zoos\, aquariums\, and science centers.
SUMMARY:Software as a Learning Context: the Case of TinkerPlots and Statist
 ical Reasoning
DTSTART:20080424T160000
DTSTAMP:20091117T205558
LOCATION:5137 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:333
SEQUENCE:0
DTEND:20080515T170000
UID:2009-11-17T20:55:58-08:00_303157444@limen.stat.ucla.edu
DESCRIPTION:The major objective in developing AEGSS is to enhance statistic
 al literacy and statistical thinking through making it possible for the stu
 dents to generate their own answers to open-ended questions rather than cho
 ose answers to multiple-choice questions. The algorithms involved in the de
 velopment of  "AEGSS" will be discussed and a demonstration of the prototyp
 e will be made. The results and overall accuracy that was obtained on a sam
 ple of essays will be discussed.
SUMMARY:Automated Essay Grading Software for Statistics (AEGSS): A Prototyp
 e
DTSTART:20080515T160000
DTSTAMP:20091117T205558
LOCATION:5137 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:336
SEQUENCE:0
DTEND:20080410T170000
UID:2009-11-17T20:55:59-08:00_207307047@limen.stat.ucla.edu
DESCRIPTION:The increased emphasis over the past decade on learning and tea
 ching in universities has been both general and discipline-based. Although 
 this sometimes causes tensions\, it is important for disciplines to be pro-
 active in analysing\, developing and proclaiming the pedagogical aspects of
  their disciplines\, including points of agreement and disagreement with th
 e general higher education literature and viewpoints. For example\, calls f
 or tertiary educators to assess what they value\, to identify learning obje
 ctives\, and to align assessment with objectives\, appear in both general a
 nd discipline-specific higher education literature emphasizing the role of 
 assessment in learning. However\, in the nexus between principles and pract
 ice in tertiary assessment in statistics and mathematics\, the variety and 
 extent of demands and pressures on assessment packages can sometimes appear
  overwhelming and even contradictory. Amidst the balancing of formative\, s
 ummative\, flexible\, continuous\, rich and authentic assessment with deman
 ds for criteria and standards-referenced assessment\, and developing generi
 c graduate capabilities such as teamwork\, problem-solving and communicatio
 n skills\, lurk the problems of over-assessment and the politics of pass ra
 tes and attrition. The many dimensions of the assessment challenge are comp
 licated in introductory statistics and mathematics courses by the diversity
  of student cohorts in which the wide range of backgrounds\, programs\, mot
 ivations and study skills need consideration in designing appropriate asses
 sment and learning packages. \n\n This presentation discusses issues\, chal
 lenges\, successful and less successful strategies in designing and impleme
 nting integrated assessment and learning packages in statistics and mathema
 tics particularly in early undergraduate years for both service and core co
 urses. The vexatious questions of plagiarism\, cooperative and group work a
 re included. Examples are given in both statistics and mathematics\, and si
 milarities and contrasts with general higher education pedagogies are highl
 ighted. \n\n Bio: Helen MacGillivray is a Professor in the Queensland Unive
 rsity of Technology's School of Mathematical Sciences\, and Director of its
  Maths Access Centre. She has taught statistics and lead statistics teachin
 g across all levels\, class sizes and many disciplines. She has written or 
 presented over 30 national and international papers in learning and teachin
 g\, and held more than 10 national or university teaching grants\, most rec
 ently a National Leadership Award and a National Senior Fellowship. She has
  also played key roles over 15 years in school syllabi\, resource developme
 nt and teacher support across all levels of schooling. \n\n Helen was the f
 irst female President\, and the first female Honorary Life Member\, of the 
 Statistical Society of Australia Inc (SSAI). She has also been President of
  the Australian Mathematical Sciences Council and is now president-elect of
  the IASE. She is currently chair of the IASE strand of the 2009 Session of
  the International Statistics Institute\, and scientific coordinator of the
  IASE's 8th International Conference on Teaching Statistics\, 2010\, and is
  Australian representative on the editorial board of the journal 'Teaching 
 Statistics'. Her current statistical research interests are in the developm
 ent and application of new distributional families of particular interest i
 n the financial world.
SUMMARY:Roles of Assessment in Learning of Statistics and Mathematics
DTSTART:20080410T160000
DTSTAMP:20091117T205559
LOCATION:5137 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:337
SEQUENCE:0
DTEND:20070601T170000
UID:2009-11-17T20:55:59-08:00_191194917@limen.stat.ucla.edu
DESCRIPTION:We will discuss the new SOCR developments\, tools and activitie
 s designed in the past year. This includes the new interactive SOCR resourc
 e viewer\, activities\, distributions\, charts\, etc. In addition\, we'll d
 iscuss the future SOCR expansions\, integration with UCLA undergraduate sta
 tistics curriculum and Moodle. Finally\, we'll have a hands-on training ses
 sion on how to use and expand the SOCR resources.
SUMMARY:The State of the Statistics Online Computational Resources
DTSTART:20070601T160000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:332
SEQUENCE:0
DTEND:20080529T170000
UID:2009-11-17T20:55:59-08:00_373041692@limen.stat.ucla.edu
DESCRIPTION:I will describe my experience writing and teaching from SticiGu
 i\, an online set of materials  for teaching Statistics.  These materials a
 re  comprised of 186 XHTML files containing about 108\,000 lines of XHTML  
 and JavaScript\,  65 Java classes containing about 16\,000 lines of code\, 
 16 JavaScript libraries containing about 5\,000 lines of code\, 34 data fil
 es containing about 5\,000 records\, a cascading style sheet with about 400
  lines\, and a handful of .jpg and .gif files. I use the materials to teach
  introductory classes\, including Berkeley's first online course. \n\n Usin
 g XHTML with Java\, JavaScript and CSS allowed me to make the content dynam
 ic: many examples and exercises change whenever the page is reloaded\, so s
 tudents can get unlimited practice at certain kinds of problems. Each stude
 nt gets a different version of each assignment\, but can see the solutions 
 to his/her version after the due date. Automation makes it easy to use mast
 ery-based assessment: students can submit each assignment up to 5 times.  O
 nly the last submission counts.  A student has to get a score of 85% or hig
 her to ""pass"" the assignment\, with a bonus for scoring 100%.  This helps
  assignments function better as learning tools instead of just yardsticks.
SUMMARY:Writing and Teaching for SticiGui\, an Online Set of Materials for 
 Teaching Statistics
DTSTART:20080529T160000
DTSTAMP:20091117T205559
LOCATION:5137 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:339
SEQUENCE:0
DTEND:20070511T170000
UID:2009-11-17T20:55:59-08:00_117148084@limen.stat.ucla.edu
DESCRIPTION:We started working on the statistics ""blended instruction"" ca
 ses study that has been funded by the College of Letters and Science in 200
 4. With the help of the OID our department chose a ""collaborative learning
  system"" called Moodle.  Jose Garcia and the Graduate Student Researchers 
 funded by this project have conducted extensive work on how to use Moodle s
 uccessfully and effectively for teaching multiple sections of classes with 
 high enrollment (100 or more students).  We have been using Moodle for teac
 hing the restructured Statistics 10 since January 2005 and presently Moodle
  is being used in as many as 22 lower division\, upper division\, and gradu
 ate courses in our department. \n\n The areas in which we have successfully
  used Moodle include: 1) Maintaining a responsive system that is available 
 twenty four hours a day so that the students can access it successfully and
  productively at any given time\, 2) updating Moodle successfully from vers
 ion1.4 in 2005 to 1.7 today\, 3) development of an automated test bank whic
 h includes one thousand multiple-choice questions most of which are written
  at the upper thinking level\, 4) Organizing the test bank so that each ins
 tructor has access to the common pool of items as well as the items in thei
 r own personal files that are not open to others\, 5) Classification of the
  items according to the major statistical concepts and strategies\, 6) Edit
 ing the items in the test bank and ascertaining that they are statistically
  sound\, 7) using Moodle to calculate the item difficulty (percentage of co
 rrect answers) of the items that have been used in the quizzes since 2005\,
   and 8) using Moodle to include more graphics in the quizzes. \n\n The cha
 llenges that remain include: 1) Using Moodle for other types of questions i
 ncluding short answers\, randomly generated questions\, and questions with 
 numerical answers\, 2)  Creating the possibility of developing parallel qui
 zzes and exams by developing the ability to post the difficulty of each ite
 m next to its title\, 2) Using Moodle for grading open-ended questions.  At
  this point we are using Moodle to have the students respond to open-ended 
 questions. But\, these responses need to be graded manually. Our objective 
 is to select and train an automated essay grading software (AEG) that would
  help us in this regard.  Based on Garcia's suggestion we need to work towa
 rd writing our own module to make this happen. In the mean time in order to
  train the software of our choice to automatically grade short open-ended q
 uestions\, we need to work on developing open-ended questions with the rele
 vant rubrics. We have already started working on this and we hope to accomp
 lish this goal by the next couple of years.
SUMMARY:Moodle: Questions Answered so Far\, Challenges Left
DTSTART:20070511T160000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:340
SEQUENCE:0
DTEND:20070504T170000
UID:2009-11-17T20:55:59-08:00_110582519@limen.stat.ucla.edu
DESCRIPTION:In the fall semester 2000 we implemented a restructured intro c
 ourse in statistics. The 'old' course was 'traditional' (in terms of pre-20
 00 instructional models). The new course transfers responsibility for learn
 ing to the student. It has 3/4 sections (60-80 students/section) that meet 
 twice a week in computer labs and twice in lecture (240-320 students).   St
 udents are given reading assignments and homework to do prior to class.  Th
 ey are given 'Readiness Assessment Quizzes' in labs on 'modules of topics'\
 , some that have been discussed in class and others not covered previously.
  The two large meetings are devoted to overviews of topics and/or small and
  large group activities.  The labs focus on individual and group work on ac
 tivities designed to enhance skills in analyzing and interpreting data. The
  impact on enrollments and instructional costs has been immense.  Some issu
 es I will address include a) value of lecturing\, b) frequent assessment\, 
 and c) value of group projects on grades.  I will discuss our experience wi
 th this revamped course and an assessment of student performance.
SUMMARY:Description\, Assessment\, Conclusions\, and 'Marijuana' (?)
DTSTART:20070504T160000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:341
SEQUENCE:0
DTEND:20070420T170000
UID:2009-11-17T20:55:59-08:00_866152182@limen.stat.ucla.edu
DESCRIPTION:When we added a lab component to Stats 10\, we hoped that it wo
 uld increase student understanding and allow students to explore larger dat
 a sets in context.  In our session we consider several concerns: \n\n \n\n 
 \n\n \n\n \n\n * How do the Fathom labs complement the instruction in Stati
 stics 10? \n\n * How do Statistics 10 students react to the Fathom labs? \n
 \n * What do two specific Fathom labs ask students to do? \n\n * How can we
  improve the lab component in Statistics 10?
SUMMARY:Using Fathom Software in Stats 10 Labs
DTSTART:20070420T160000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:342
SEQUENCE:0
DTEND:20061121T160000
UID:2009-11-17T20:55:59-08:00_392279273@limen.stat.ucla.edu
DESCRIPTION:The aim of my talk is to present you the results achieved and t
 he experience gained in designing an implementing e-status (http://ka.upc.e
 s)\, a web-based tool to improve students performance in statistics and the
  formal randomized evaluation of its effects. \n\n This talk is an updated 
 version of the work developed together with my colleagues\, presented in IC
 OTS7  (Gonzalez et al.\, 2006) and the recently published paper in CAEE  (G
 onzalez and Munoz\, 2006) \n\n e-status is a tool developed mixing together
  the main working lines in stats education: Learning by practicing and Prob
 lem Solving\, to which the feedback knowledge of students improvement has b
 een added.   The tool is complementary to classical learning materials used
 : practical sessions at the computer lab and lecture-based instruction. \n\
 n This talk is co-sponsored by the UCLA SOCR project.
SUMMARY:Formal Assessment of a Web-based Tool Designed to Improve Student P
 erformance in Statistics
DTSTART:20061121T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:338
SEQUENCE:0
DTEND:20070525T170000
UID:2009-11-17T20:55:59-08:00_699856205@limen.stat.ucla.edu
DESCRIPTION:This talk is addressed to teachers of Mathematics across the sc
 hool curriculum (K-12) that will introduce in their classes reasoning with 
 data and chance according to the most recent guidelines of the NCTM\, GAISE
  and College Board Guidelines in 2006. It will show how to make the best us
 e of what is already there in the CensusAtSchool International Project (htt
 p://www.censusatschool.ntu.ac.uk/)  to prepare activities for the classroom
   and to involve students in thinking and reasoning about data and chance w
 ith minimum effort while learning about other children in the world. The ex
 perience of the 5 participating countries (Australia\, New Zealand\, Canada
 \, United Kingdom and South Africa) will be summarized with video clips. \n
 \n The talk will also explain how teachers in California can get their clas
 s involved in the new phase of this International project at no cost\, if t
 hey wish. \n\n The National Council of Teacher of Mathematics (2006) includ
 es CensusAtSchool as an example of one of the ways to introduce the new gui
 delines across the curriculum\, and that article will be handed out with pe
 rmission of the authors at the seminar.  Other promotional  material from C
 ensusAtSchool in the participating countries will be distributed at the sem
 inar. \n\n The talk will probably be videotaped.
SUMMARY:The CensusAtSchool International Project (http://www.censusatschool
 .ntu.ac.uk/)
DTSTART:20070525T160000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:343
SEQUENCE:0
DTEND:20060601T160000
UID:2009-11-17T20:55:59-08:00_998406843@limen.stat.ucla.edu
DESCRIPTION:In this presentation we will discuss: \n\n 1. The preliminary f
 indings of an exploratory study on teaching statistics in the community col
 leges within the Los Angeles County. This will include examination of typic
 al syllabi for introductory statistics\, statistic textbook used by 80% - 9
 0% of the community colleges\, and student assessment. \n\n 2. Guidelines p
 rovided by the American Statistical Association for teaching introductory s
 tatistics (GAISE) and the extent to which the community college instructors
  are familiar with them. \n\n 3. Summary of one-on-one interviews conducted
  with a number of community college statistics instructors within the Los A
 ngeles County.
SUMMARY:An Overview of Teaching Introductory Statistics in Community Colleg
 es
DTSTART:20060601T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:344
SEQUENCE:0
DTEND:20060525T160000
UID:2009-11-17T20:55:59-08:00_823350293@limen.stat.ucla.edu
DESCRIPTION:With increasing college enrollments and availability of technol
 ogy\, many campuses are investigating the use of online instruction. In the
  Fall of 2001\, as part of a larger study at the University of California\,
  Davis\, we conducted an experiment to compare a "hybrid" offering with a t
 raditional offering of our large introductory statistics course.  For the h
 ybrid offering the class met once a week for evaluation and an overview\, b
 ut the students were required to learn the material using web-based materia
 l (CyberStats) and a textbook. We examined differences in student performan
 ce\, student satisfaction and investment of both student and instructor tim
 e for the hybrid and traditional classes. The hybrid course was taught agai
 n in Fall 2002\, modified based on the study results\, but still wasn't qui
 te right. A more successful approach was used in Fall 2003 and 2004. This t
 alk will discuss the results of the study\, changes in the course since the
 n\, and recommendations in the context of Garfield's (1995) principles of l
 earning statistics.
SUMMARY:An Experiment Comparing Hybrid and In Class Instruction in Introduc
 tory Statistics
DTSTART:20060525T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:345
SEQUENCE:0
DTEND:20060518T160000
UID:2009-11-17T20:55:59-08:00_843103193@limen.stat.ucla.edu
DESCRIPTION:Although the Statistics Education community has advocated using
  real data to teach introductory statistics for quite some time\, often the
 se data sets are not recognizably real to statisticians since the students'
  limited experience with "real" statistical software and data management te
 chniques precludes the use of truly messy data. But grappling with messy an
 d complex data sets is important for teaching Statistical Thinking (broadly
  defined as "thinking like a statistician") and is appropriate for an intro
 ductory statistics course. We describe our experience collecting rich data 
 sets and developing computer lab assignments using STATA to teach statistic
 al thinking to first-year university students using these data sets. Collec
 ting useable\, real\, data sets turns out to be fairly difficult for severa
 l reasons\, and teaching data management and analysis without resorting to 
 rote-based rules is quite challenging.
SUMMARY:Towards Statistical Thinking: Making Real Data Real
DTSTART:20060518T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:346
SEQUENCE:0
DTEND:20060511T160000
UID:2009-11-17T20:55:59-08:00_640658550@limen.stat.ucla.edu
DESCRIPTION:In May 2005\, ASA endorsed guidelines for instruction in statis
 tics education at the PreK-12 level and the college introductory course. Th
 ese guidelines were developed by a group of leading statistics and mathemat
 ics educators. The foundation for the PreK-12 framework rests on the NCTM S
 tandards. The PreK-12 document provides guidance on the content that should
  be taught in the elementary\, middle and high school grades\, focusing on 
 a connected curriculum that will allow a high school graduate to have a wor
 king knowledge and appreciation for the basic ideas of statistics. The coll
 ege document focuses on recommendations for the introductory course that pr
 omote conceptual understanding and the attainment of statistical literacy a
 nd thinking. Both documents have the same ultimate goal: developing a stati
 stically literate citizen. \n\n An overview of the documents along with sho
 rt examples of moving statistical concepts across developments levels will 
 be presented. Issues associated with the implementation of the recommendati
 ons will also be discussed. \n\n Bio-sketch: Christine Franklin is a Lectur
 er and Honors Professor in the Department of Statistics at the University o
 f Georgia. She has been actively involved at the national level with promot
 ing statistical education at the Pre K-12 level since the early 1990's. Her
  involvement with the AP Statistics program includes preparing teachers sin
 ce 1995 to teach AP Statistics and serving as a table and question leader a
 t the AP Statistics readings. She also conducts College Board workshops and
  in 1998\, began teaching a new course at UGA that she designed for seconda
 ry math teachers. She has also developed a new master's level course in pro
 bability and statistics for 6-8 teachers and is currently developing a new 
 course for Pre K-5 teachers. \n\n Chris has served on the ASA Advisory Comm
 ittee for Teacher Enhancement since 2000 and served as chair of the steerin
 g committee that planned the ASA sponsored inaugural conference in statisti
 cs for teacher educators (TEAMS) that was held in October 2003. She is an a
 ssociate editor for the Journal of Statistics Education. She chaired the AS
 A project\, GAISE\, for developing Pre K-12 guidelines in statistical educa
 tion. She is the 2006 chair of the ASA Section on Statistical Education and
  is a Fellow of ASA. \n\n Chris has been honored with numerous teaching and
  advising awards at the University of Georgia. Chris has also written numer
 ous publications for textbooks and educational journals. A recent project w
 as writing for the NCTM Navigation Series where she is one of four authors 
 for the 9-12 Data Analysis book. She is the co-author along with Alan Agres
 ti of an introductory college level statistics textbook for Prentice Hall\,
  published in January 2006. \n\n Chris served as an advisor to the GA mathe
 matics committee that recently revised the Georgia Pre K-12 mathematics sta
 ndards. The GAISE Framework was an instrumental part with infusing more dat
 a analysis into the new GA Performance Standards. \n\n Chris has two boys\,
  Corey who graduates from high school in May\, and Cody\, a rising 6th grad
 er. She loves to run\, hike and backpack\, play the piano\, play softball\,
  read\, and attend basketball and baseball games. She hates to cook but lov
 es to eat. But most of all\, she enjoys spending time with her husband and 
 boys.
SUMMARY:American Statistical Association (ASA) Guidelines for Instruction i
 n Statistics Education  (GAISE) from PreK-12 and the College Intro Course: 
 What does this mean for the Pre K-12 mathematics curriculum and AP Statisti
 cs?
DTSTART:20060511T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:348
SEQUENCE:0
DTEND:20060420T160000
UID:2009-11-17T20:55:59-08:00_310584395@limen.stat.ucla.edu
DESCRIPTION:Teachers of statistics make frequent decisions about how to bes
 t use technology and how to best teach technology.  The UCLA Department of 
 Statistics is about to launch the first journal to publish research intende
 d to help the statistics education make data-based decisions on technology 
 choices.  Currently\, the leading U.S. journal of statistics education is t
 he ASA's  Journal of Statistics Education (JSE).  We therefore examine the 
 JSE to see how it addresses the "technology issue".  We apply a multi-dimen
 sional scaling technique to create a "map" of the content of the JSE\, and 
 in this talk will explore this map and its variations to better understand 
 how our new journal might address these issues.
SUMMARY:Technology\, Statistics\, and Teaching: What Do Our Journals Tell U
 s?
DTSTART:20060420T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:349
SEQUENCE:0
DTEND:20060406T160000
UID:2009-11-17T20:55:59-08:00_597983814@limen.stat.ucla.edu
DESCRIPTION:I will describe the new developments\, tools and online materia
 ls part of the UCLA SOCR resource. This includes: \n\n * Resource Organizat
 ion \n\n * Features and Functionalities \n\n * Pedagogical utilization \n\n
  * SOCR Assessment \n\n * Future developments \n\n This is joint work with 
 Annie Che and Jenny Cui\, funded by NSF DUE 0442992 and NIH U54 RR021813.
SUMMARY:The SOCR Site: A Collection of Applets for Teaching Probability and
  Statistics
DTSTART:20060406T150000
DTSTAMP:20091117T205559
LOCATION:5127 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:350
SEQUENCE:0
DTEND:20050603T160000
UID:2009-11-17T20:55:59-08:00_777508942@limen.stat.ucla.edu
DESCRIPTION:This presentation consists of four parts: \n\n 1.  In the first
  part the "generative model" proposed by Merlin Wittrock will be discussed 
 as the theoretical foundation behind teaching for thinking and generation o
 f knowledge by learners. \n\n 2.  In the second part "blended instruction"\
 , implementation of "generative model" combined with technology and on-line
  quizzes\, will be discussed as a mean of teaching for thinking in "Introdu
 ctory Statistics". \n\n 3.  In the third part\, the concentration will be o
 n: \n\n * Potential strategies the instructor can follow in designing multi
 ple-choice and short-answer questions that make it possible to test for thi
 nking in "introductory statistics"\, and \n\n * Examples of multiple-choice
  and short answer questions that help the instructor to test for thinking i
 n "Introductory Statistics". \n\n In the fourth part the "challenge of teac
 hing and testing for thinking" in large "Introductory Statistics" classes o
 f 100-150 will be discussed.
SUMMARY:Teaching and Testing for Thinking in Introductory Statistics
DTSTART:20050603T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:352
SEQUENCE:0
DTEND:20050506T160000
UID:2009-11-17T20:55:59-08:00_955222705@limen.stat.ucla.edu
DESCRIPTION:While many statistical consulting projects can be kept straight
 forward\, drawing upon well polished skills of a good consultant\, others p
 rojects are inherently more difficult.   In some cases\, the needs of the c
 lient can only be met with a push toward a "second solution" that avoids ov
 ersimplifications and untenable assumptions that plague the first solution 
 approach.   A good consultant recognizes these situations\, and an ambitiou
 s consultant embraces them.  As these opportunities develop\, the relations
 hip between the client and consultant evolves from a client-server model to
  a peer-to-peer model.  When the client need matches the consultant's inter
 est\, the work can lead to some novel collaborative research.  In this talk
 \, we comment on some experience with this process\, and offers suggestions
  as to how to nurture it.  We also highlight how statistics graduate studen
 t training at UCR Has been enhanced by collaborative research in our consul
 ting center.
SUMMARY:The Transition from 'Service' to 'Collaboration' in a Statistical C
 onsulting Environment
DTSTART:20050506T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:354
SEQUENCE:0
DTEND:20050422T160000
UID:2009-11-17T20:55:59-08:00_787370248@limen.stat.ucla.edu
DESCRIPTION:Videos of classroom teaching collected as part of the Third Int
 ernational Mathematics and Science Study reveal that teaching is a cultural
  activity\, varying more across cultures than within. It is learned implici
 tly\; it is largely based on hidden cultural scripts\; it is embedded in wi
 der cultural beliefs and practices\; and it is difficult to change. Given t
 hese facts\, how can teaching be improved? In this presentation I will brie
 fly describe most recent findings from the TIMSS Video Studies of mathemati
 cs teaching in seven countries\, and discuss the implications of these find
 ings for (a) current debates about mathematics teaching and learning in sch
 ools\, and (b) teacher professional learning.
SUMMARY:The Teaching Gap: Reflections on Teaching and How to Improve It
DTSTART:20050422T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:355
SEQUENCE:0
DTEND:20040611T160000
UID:2009-11-17T20:55:59-08:00_69837778@limen.stat.ucla.edu
DESCRIPTION:Every year the UCLA Office of Instructional Development (OID) a
 ccepts proposals from approximately 15 advanced graduate students to teach 
 a seminar course of their own design.  I was selected  to teach “Making Sen
 se of Lies\, Damned Lies and Statistics” this Spring.  The main objective o
 f the course was not to teach students how to calculate statistics on their
  own or to run their own experiments\, but to develop a critical statistica
 l eye for evaluating statistical information found in mass media.  This cou
 rse also differs from standard introductory statistics courses in that OID 
 required for both writing and discussion to be the main forms of student ev
 aluation.  This presentation will discuss the approach taken to achieve the
  course objective while meeting the requirements of the OID Collegium of Un
 iversity Teaching Fellows program\, my assessment of how well the goals wer
 e met\, and what some of the student feedback has been thus far.  I will sh
 are what aspects have gone particularly well\, and what I would improve upo
 n if I were to teach a similar course in the future.
SUMMARY:Collegium Fellows Seminar:  Teaching Undergraduate Statistics Stude
 nts How to Read the Newspaper
DTSTART:20040611T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:356
SEQUENCE:0
DTEND:20040604T160000
UID:2009-11-17T20:55:59-08:00_518568597@limen.stat.ucla.edu
DESCRIPTION:In this presentation I will discuss the potentials and limitati
 ons of GIS systems as an entry point for statistical reasoning in the conte
 xt of highly-charged\, meaningful social issues. \n\n First\, I will make t
 he general case for the use of GIS and social justice issues to teach stati
 stics. \n\n Second\, I will examine a pilot study of 28 high school student
 s who engaged in a GIS project as part of a summer seminar in social resear
 ch. Analyses of pre- and post-tests\, and the students’ final PowerPoint pr
 esentations will be presented to document conceptual growth as well as the 
 way in which mathematical reasoning was limited by the combination of the G
 IS tools and rhetorical structure of the projects.
SUMMARY:Mapping Educational Injustice 50 Years After Brown v. Board of Educ
 ation: Bridges to Qualitative Shifts in Quantitative Reasoning
DTSTART:20040604T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:358
SEQUENCE:0
DTEND:20040518T160000
UID:2009-11-17T20:55:59-08:00_353634260@limen.stat.ucla.edu
DESCRIPTION:We will describe the philosophy\, design\, implementation and u
 tilization of the Statistical Online Computational Resource (SOCR) for unde
 rgraduate and graduate education. Classical Probability and statistics inst
 ruction can be significantly enhanced by employing interactive\, contempora
 ry\, computational\, visualizational and graphical tools\, in and out of cl
 ass. We have developed one such suite of tools (SOCR) that consists of virt
 ual experiments\, distribution models\, games\, statistical analysis tools 
 and  other additional models (e.g.\, curve fittings\, parameter estimates\,
  probability tables\, etc.) \n\n All of these tools are designed in a moder
 n object oriented fashion and implemented as look-alike Java applets. The S
 OCR resources may be accessed by students from any Java-enabled computer. I
  will discuss how these tools have been used in my graduate and undergradua
 te classes to demonstrate statistical concepts\, experimental properties\, 
 analytic methods and data manipulations (e.g.\, visualization). \n\n Over t
 he past 3 years many statistics faculty and students have been involved in 
 the development of the SOCR suite of tools (see online SOCR acknowledgments
 ).
SUMMARY:Statistics Online Computational Resource for Education
DTSTART:20040518T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:360
SEQUENCE:0
DTEND:20040507T160000
UID:2009-11-17T20:55:59-08:00_618413533@limen.stat.ucla.edu
DESCRIPTION:The post-calculus Introductory Statistics course for Engineers 
 and Computer Scientists is usually taught with textbooks that have very few
  examples for the computer science major\, such as analysis of computer per
 formance and algorithms\, web browsing behavior of users\, search engines\,
  the Internet network\, spam filtering\, who uses the Internet\, and other.
  In this talk\, I will introduce some activities that I have prepared speci
 fically to fill that gap. Most of these activities involve data analysis. I
  will also present an assessment of how some of the activities worked out i
 n a class of 28 students where 80 percent of the students were either compu
 ter science majors alone or both computer science/engineering majors and th
 e rest were math\, applied math or econ/math students. I will describe whic
 h students chose which projects\, and I will illustrate their work with a s
 ample of their papers.
SUMMARY:Assessing New Activities for Computer Science Majors in a Post-calc
 ulus Introductory Statistics Course
DTSTART:20040507T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:361
SEQUENCE:0
DTEND:20040416T160000
UID:2009-11-17T20:55:59-08:00_4892881@limen.stat.ucla.edu
DESCRIPTION:I'll describe a new professional development program for second
 ary mathematics teachers who are preparing to teach statistics\, and addres
 s what we have learned in our efforts to design a course that has a signifi
 cant online component and that is relevant and useful from a teacher’s pers
 pective.  The ways in which our online environment incorporates group work\
 , self-study\, exploration of concepts\, and assessments are described.  Th
 e challenges associated with delivering the necessary content while at the 
 same time recognizing the practical time constraints of adult students who 
 are themselves teaching full-time are also discussed.
SUMMARY:Preparing Secondary Mathematics Teachers to Teach Statistics
DTSTART:20040416T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:363
SEQUENCE:0
DTEND:20030502T160000
UID:2009-11-17T20:55:59-08:00_841569518@limen.stat.ucla.edu
DESCRIPTION:In this talk\, Dr. Brian Jersky\, Chair of the Mathematics Deap
 rtment at Sonoma State University (SSU) and co-director of the SSU Statisti
 cal Consulting Center\, will present an overview of how a new statistical c
 onsulting center\, together with a new statistical consulting class at SSU\
 , have resulted in increased opportunities for outreach into the surroundin
 g community for both students and faculty\, as well as increasing the numbe
 r of majors in the Statistics track of SSU's Mathematics major. Examples of
  projects that have worked\, and some that didn't\, will be presented. Ther
 e will be time for questions during and after the talk.
SUMMARY:Statistical Consulting: A Nexus for Community-based and In-class Le
 arning
DTSTART:20030502T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:364
SEQUENCE:0
DTEND:20030418T160000
UID:2009-11-17T20:55:59-08:00_792805561@limen.stat.ucla.edu
DESCRIPTION:Most statistics courses integrate technology into the course. B
 ut often they just tack the technology onto the old syllabus. We propose a 
 new alternative. We characterize what Statistics students need to know in t
 hree broad steps: Think\, Show\, and Tell. Think comprises identifying the 
 variables\, selecting methods\, and checking conditions. The Show step is f
 inding the numerical answers. Tell is the all-important (and often ignored)
  step of explaining the findings and drawing conclusions. Many traditional 
 Statistics courses are\, in this scheme\, “just Show”. But the Show step is
  exactly what technology does well and what practicing Statisticians rely o
 n technology to do. When students rely on their technology for the calculat
 ions\, there is more time to concentrate on the Think and Tell steps. This 
 approach swings the emphasis of the introductory statistics course toward s
 tatistical thinking and understanding and away from calculating statistics.
  We can also select formulas for the Show step that emphasize understanding
 \, even if they would not have been the first choice for hand computing. We
  will report insights and practical solutions arising from our work on a ne
 w introductory textbook that follows this reasoning. \n\n Joint work with R
 ichard D. DeVeaux\, Williams College
SUMMARY:Statistics: Ready\, Tech\, Go\; If Technology Has Revolutionized th
 e Teaching of Statistics\, Why are We Still Teaching Essentially the Same C
 ourse?
DTSTART:20030418T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:368
SEQUENCE:0
DTEND:19971006T160000
UID:2009-11-17T20:55:59-08:00_166787569@limen.stat.ucla.edu
DESCRIPTION:After a brief survey of classical models for stocks evolution a
 nd the related option pricing problem\, I will concentrate on T.W.Epps' mod
 el (Comm. Statist. 1996) for stocks behaviour\, show that it is actually a 
 branching process in random environment and use results from Dion & Esty (A
 nn. Statist.\, 1977) to estimate the main parameters of the system and get 
 a heuristic goodness of fit test. Its usefulness in modeling daily closing 
 prices at the Stock Exchange\, 1962-1987\, will be discussed.
SUMMARY:Stocks and Options: from Brownian Motions to Branching Processes
DTSTART:19971006T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:365
SEQUENCE:0
DTEND:20030403T160000
UID:2009-11-17T20:55:59-08:00_169230587@limen.stat.ucla.edu
DESCRIPTION:This presentation will describe the emergence of statistics edu
 cation as a unique discipline over the past thirty years. The research lite
 rature supporting this discipline will be summarized and implications of th
 is literature for teaching and assessing students will be suggested. New de
 velopments and important projects in statistics education will be shared.
SUMMARY:Statistics Education: An Emerging Discipline
DTSTART:20030403T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:366
SEQUENCE:0
DTEND:19971022T170000
UID:2009-11-17T20:55:59-08:00_110319031@limen.stat.ucla.edu
DESCRIPTION:A method for estimating the coalescence time (time since the mo
 st recent common ancestor) of a sample of intraspecies DNA sequences utiliz
 ing full data from the sample is proposed. The sequences evolve according t
 o the finitely-many-sites model for base substitution with several possible
  generator matrices. The coalescence time is estimated on the basis of a sa
 mple from a posterior distribution of genealogical trees given the observed
  nucleotide sequences. To generate this sample\, a Markov chain on the spac
 e of genealogical trees is constructed based on the Metropolis-Hastings alg
 orithm. The strategy for updating the current genealogical tree (the state 
 of the chain) is discussed in detail. Two other updating algorithms related
  to the proposed method\, by Kuhner et al. (1995)\, and by Pearl et al.(199
 6)\, are also discussed. The approach can be extended to allow for uncertai
 nty about the value of the mutation rate\, as well as for variable populati
 on size and variable mutation rate within the DNA region of interest.
SUMMARY:Estimating the Coalescence Time from DNA Sequence Data Using Markov
  Chain Monte Carlo Techniques
DTSTART:19971022T160000
DTSTAMP:20091117T205559
LOCATION:6221 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:369
SEQUENCE:0
DTEND:19971001T160000
UID:2009-11-17T20:55:59-08:00_421182945@limen.stat.ucla.edu
DESCRIPTION:In this talk\, we try to investigate sinusoidal transformations
  of stochastic processes.  These models arise from communications.  However
 \, they may also be used for speech coding. First of all\, we study the spe
 ctrum of these models. Then we find the exact or approximate ARMA forms for
  these models.  Their limiting distribution and other properties are obtain
 ed.  Further\, we develop a recursive algorithm to calculate the Cramer-Rao
  lower bound for tracking the stochastic process inside the sinusoidal func
 tion (Demodulation) and discuss several efficient filtering methods in the 
 presence of noise.  Finally\, we try to apply these models for speech codin
 g.
SUMMARY:Frequency Modulation Models with Applications to Communications and
  Speech Coding
DTSTART:19971001T150000
DTSTAMP:20091117T205559
LOCATION:5117 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:370
SEQUENCE:0
DTEND:19970512T160000
UID:2009-11-17T20:55:59-08:00_85691294@limen.stat.ucla.edu
DESCRIPTION:The EM algorithm is a popular method for computing maximum like
 lihood estimates. One of its primary drawbacks is that it does not produce 
 standard errors as a by-product. In this paper we consider obtaining standa
 rd errors by numerical differentiation. Two approaches are considered. The 
 first differentiates the Fisher score vector to give the Hessian of the log
 -likelihood. The second differentiates the EM operator and uses an identity
  that relates it to the Hessian of the log-likelihood. The well-known SEM a
 lgorithm uses the second approach. We consider three additional algorithms\
 , one that uses the first approach and two that use the second. \n\n We eva
 luate the complexity and precision of these three and SEM in 10 examples. T
 he first is a single parameter example used to give insight. The others are
  three examples in each of three areas of EM application: Poisson mixture m
 odels\, estimation of covariance from incomplete data\, and confirmatory fa
 ctor analysis. The examples show there are algorithms that are significantl
 y simpler and more accurate than SEM. Hopefully their simplicity will incre
 ase the availability of standard error estimates in EM applications. \n\n I
 t is shown that as previously conjectured  a symmetry diagnostic can accura
 tely estimate errors arising from numerical differentiation. A number of is
 sues related to the speed of the EM algorithm and algorithms that use the t
 he second approach and in particular SEM are identified.
SUMMARY:Standard Errors for EM Estimation
DTSTART:19970512T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:372
SEQUENCE:0
DTEND:19970428T160000
UID:2009-11-17T20:55:59-08:00_28165816@limen.stat.ucla.edu
DESCRIPTION:The sample covariance matrix is singular when dimension exceeds
  sample size. Improved estimators in the literature are either not defined 
 or not invertible in this case. In addition\, they require normality and ar
 e only shown to dominate the sample covariance matrix\, not to be optimal w
 ithin a reasonable class. \n\n This paper finds the shrinkage estimator wit
 h minimum quadratic risk in the class of linear combinations of the identit
 y and the sample covariance matrix\, asymptotically as the number of variab
 les and the number of observations go to infinity together. This estimator 
 remains well-justified and invertible even if variables outnumber observati
 ons. It does not require normality. Its simple explicit formula is easy to 
 compute and to interpret. Extensive Monte-Carlo confirm that the asymptotic
  results hold well in finite sample.
SUMMARY:A Well-conditioned Estimator for Large Dimensional Covariance Matri
 ces
DTSTART:19970428T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:373
SEQUENCE:0
DTEND:19970421T160000
UID:2009-11-17T20:55:59-08:00_143048705@limen.stat.ucla.edu
DESCRIPTION:Modeling a DNA sequence as a sequence of letters generated by a
  stationary Markov chain\, we derive a Poisson process approximation for th
 e occurrences of clumps of multiple words\, and a compound Poisson process 
 approximation for the number of occurrences of multiple words. Using the Ch
 en-Stein method we also give a bound on the error in the approximation. The
  result is applied to approximate the number of occurrences of certain stem
 -loop motifs. \n\n This is joint work with S.Schbath\, INRA\, France
SUMMARY:Compound Poisson and Poisson Process Approximations for Occurrences
  of Multiple Words\, and Application to Stem-loop Motifs in DNA Sequences
DTSTART:19970421T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:376
SEQUENCE:0
DTEND:19970331T160000
UID:2009-11-17T20:55:59-08:00_399089575@limen.stat.ucla.edu
DESCRIPTION:We introduce and discuss special notions of likelihood factoriz
 ations. Such a factorization  may be either parameter-based and therefore p
 ermit a split of parameters into variation independent components or they m
 ay correspond to a given concentration graph\, i.e. an undirected Markov gr
 aph for  several random variables in which each edge represents a statement
  about the pairwise conditional association given all remaining variables. 
 \n\n If a factorization is both parameter-based and a concentration graph f
 actorization  then a joint likelihood analysis can be split into separate a
 nalyses each involving a reduced set of component variables and parameters.
  \n\n Conditions on the concentration graph which help to decide on this de
 sirable feature involve the notions of complete separators and of prime gra
 phs. \n\n First\,  the available results for joint Gaussian distributions a
 nd for contingency table analyses are restated\, then two parametric famili
 es of distributions for mixed discrete and continuous variables are discuss
 ed leading to  conditional Gaussian models and to partially dichotomized Ga
 ussian models. The conditons on the graphs show marked differences between 
 these two distributional families.
SUMMARY:Can We Read Off Undirected Markov Graphs Whether a Multivariate Est
 imation Problem can be Split into Separate Smaller and Simpler Problems
DTSTART:19970331T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:378
SEQUENCE:0
DTEND:19970317T160000
UID:2009-11-17T20:55:59-08:00_316056316@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Linear Processes\, Long-range Dependence and Asymptotic Expansions
DTSTART:19970317T150000
DTSTAMP:20091117T205559
LOCATION:5117 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:375
SEQUENCE:0
DTEND:19970407T160000
UID:2009-11-17T20:55:59-08:00_3916207@limen.stat.ucla.edu
DESCRIPTION:Data mining is a necessary crime in statistics.  A good modelin
 g process almost always involves careful data exploration\, model selection
 \, and diagnostics. However\, most of the standard statistical theory is va
 lid only if the model is formed 'a priori'. Ignoring the data mining proces
 s often leads to seriously over-optimistic inferences. Therefore there is a
 n inconsistency problem between the standard theory and practice. \n\n We o
 ffer a systematic framework under which complex data mining procedures can 
 be analyzed in the same way as the classical linear models.  This framework
  centers around the concept of generalized degrees of freedom (GDF) which c
 haracterizes the cost of a general modeling process.  Using this framework\
 , many difficult problems can be solved easily. For example\, we can now me
 asure the number of observations cost in a variable selection process.  Dif
 ferent modeling procedures\, such as a tree-based regression and a projecti
 on pursuit regression\, can be compared on the basis of their residual sums
  of squares and the degrees of freedom they cost. \n\n We will apply the pr
 oposed framework to the problems of model selection.  A unbiased estimates 
 of the error variance based on the selected model will be obtained. Further
  applications result in an unbiased dimension selection procedure and an ad
 aptive model selection that uses a data-driven penalty. Both procedures are
  shown to be substantially superior to conventional model selection procedu
 res.
SUMMARY:On Measuring and Correcting the Effects of Data Mining and Model Se
 lection
DTSTART:19970407T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:377
SEQUENCE:0
DTEND:19970318T160000
UID:2009-11-17T20:55:59-08:00_803436043@limen.stat.ucla.edu
DESCRIPTION:Genetic mapping consist of ordering sets of loci and estimating
  distances between pairs of loci.  One commonly used distance is the geneti
 c distance\, defined as the expected number of crossovers occurring between
  the loci.  Crossing over occurs during meiosis\, the cellular division res
 ulting in sex cells.  Unfortunately\, it is often the case that the number 
 of crossovers cannot be observed.  Rather\, the data consist of observation
 s of recombination (an odd number of crossovers occurring between loci). \n
 \n Estimates of genetic distance have typically been made ignoring interfer
 ence (dependence) between recombinations as well as the possiblity of typin
 g errors (i.e. misspecifying the likelihood).  However\, both errors and in
 terference impact estimates of genetic distance.  After providing relevant 
 biological background\, I  examine the impact of genotyping errors on dista
 nce estimates using the chi-square model of recombination\, which allows fo
 r interference.  Results on the relationships between true distance\, error
  rate\, and degree of interference will be presented.
SUMMARY:Effects of Genotyping Errors and Interference on Estimates of Genet
 ic Distance
DTSTART:19970318T150000
DTSTAMP:20091117T205559
LOCATION:5117 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:379
SEQUENCE:0
DTEND:19970314T150000
UID:2009-11-17T20:55:59-08:00_561060848@limen.stat.ucla.edu
DESCRIPTION:The estimation of normalizing constants for a family of distrib
 utions is a recurrent theme in computational statistics.  After a brief des
 cription of a number of applications of interest\, including likelihood cal
 culation in a genetic linkage problem\, I will proceed to examine two diffe
 rent approaches: bridge sampling (Meng and Wong\, 1996) and maximum profile
  likelihood (inverse logistic regression\, Geyer\, 1993\, and Kong\, 1996) 
 with an emphasis on the extension to the case of sampling from more than tw
 o distributions and on the equivalence of the two methods. \n\n In the end\
 , I will propose new estimation procedures based on tractable transformatio
 ns of the underlying distributions of the samples\, designed for dealing wi
 th multimodal continuous densities.
SUMMARY:Efficient Estimation of Normalizing Constants from Markov Chain Mon
 te Carlo Draws
DTSTART:19970314T140000
DTSTAMP:20091117T205559
LOCATION:6221 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:383
SEQUENCE:0
DTEND:19970224T160000
UID:2009-11-17T20:55:59-08:00_458575781@limen.stat.ucla.edu
DESCRIPTION:If an experimenter wants to determine the degree of a polynomia
 l regression on the basis of a sample of observations\, Anderson (1962) sho
 wed that the following method is optimal. Starting with the highest (specif
 ied) degree the procedure is to test in sequence whether the coefficients a
 re 0.  In this paper optimal designs for Anderson's procedure are determine
 d explicitly.  The optimal design maximizes the minimum power of a given se
 t of alternatives.
SUMMARY:Optimal Designs for Identifying the Degree of a Polynomial Regressi
 on
DTSTART:19970224T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:380
SEQUENCE:0
DTEND:19970310T160000
UID:2009-11-17T20:55:59-08:00_935252413@limen.stat.ucla.edu
DESCRIPTION:In this talk\, a general theory for the construction of confide
 nce intervals or regions in the context of nonstationary dependent data wil
 l be presented. The basic idea is to approximate the sampling distribution 
 of a statistic based on the values of the statistic computed over smaller s
 ubsets of the data. We present a general asymptotic validity result under m
 inimal conditions. Some simulation studies shed light on the small sample p
 erformance of the method. As an application\, we address the problem whethe
 r stock returns can be predicted from dividend yields.
SUMMARY:Subsampling for Nonstationary Time Series
DTSTART:19970310T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:382
SEQUENCE:0
DTEND:19970227T160000
UID:2009-11-17T20:55:59-08:00_388381080@limen.stat.ucla.edu
DESCRIPTION:In this seminar\, we will discuss several strategies for creati
 ng digital atlases of the human brain. In particular\, we describe our rece
 nt progress towards creating a probabilistic brain atlas based on high-dime
 nsional random fluid transformations. An array of computational tools are u
 sed to analyze a reference image archive of scans from several patient popu
 lations. These tools are then used to detect and quantify anatomic abnormal
 ities in brain scans of new subjects. \n\n Striking variations exist\, acro
 ss individuals\, in the internal and external geometry of the brain. In the
  past\, quantifying deviations from normal anatomy and comparing functional
  data from different subjects or patient subpopulations have been difficult
  because cortical topography and the internal geometry of the brain vary so
  greatly. In this seminar\, we describe the design\, implementation and pre
 liminary results of a technique for creating a comprehensive probabilistic 
 atlas of the human brain based on high-dimensional fluid transformations. H
 igh-dimensional warping algorithms fluidly deform one subject's anatomy int
 o structural correspondence with another\, and enable the transfer of 3D fu
 nctional data between subjects or onto a single anatomic template\, for sub
 sequent comparison or integration. The goal of the probabilistic atlas is t
 o detect and quantify subtle and distributed patterns of deviation from nor
 mal anatomy\, in a 3D brain image from any given subject. The algorithm ana
 lyzes a reference population of normal scans\, and automatically generates 
 color-coded probability maps of the anatomy of new subjects.
SUMMARY:Mathematical/Computational Strategies for Analyzing Human Brain
DTSTART:19970227T150000
DTSTAMP:20091117T205559
LOCATION:6221 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:384
SEQUENCE:0
DTEND:19970220T160000
UID:2009-11-17T20:55:59-08:00_175211681@limen.stat.ucla.edu
DESCRIPTION:For signal and image classification and discrimination\, in par
 ticular\, in medical or geophysical diagnostics and military applications\,
  extracting relevant features and reducing the dimensionality of measured d
 ata is of vital importance. As an attempt to automate the feature extractio
 n procedure and to understand what the critical features for these problems
 \, we developed the so-called local discriminant basis (LDB) method which r
 apidly selects an orthonormal basis suitable for signal/image classificatio
 n problems from a large collection of orthonormal bases (e.g.\, wavelet pac
 kets and local trigonometric bases). Once the LDB is selected\, a small num
 ber of most significant coordinates (features) are fed into a traditional c
 lassifier such as linear discriminant analysis or decision trees. The perfo
 rmance of these statistical methods is enhanced since the LDB method reduce
 s the dimensionality of the problems without losing important information f
 or classification. Moreover\, since the basis functions well-localized in t
 he time-frequency plane are used as feature extractors\, interpretation of 
 the classification results becomes easier and more intuitive than using the
  conventional methods. \n\n In this talk\, after briefly reviewing our tool
 kit (i.e.\, wavelet packets and local trigonometric functions)\, we describ
 e the original LDB method (which maximizes certain "distances" among time-f
 requency energy distributions of signal classes) as well as its recent prog
 ress (which maximizes the "distances" among empirical probability densities
  of basis coordinates). We also show an application of these techniques to 
 a real geophysical problem of classifying acoustic waveforms propagated in 
 a borehole according to the lithology of geological formations. This talk w
 ill conclude with future directions of the adapted feature extraction techn
 ology\, in particular\, applications to cluster analysis of high dimensiona
 l datasets which turn out to have a deep connection with geometric analysis
 . \n\n This is a joint work with Prof. Ronald R. Coifman at Yale University
 .
SUMMARY:Adapted Feature Extraction and Its Applications
DTSTART:19970220T150000
DTSTAMP:20091117T205559
LOCATION:6221 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:386
SEQUENCE:0
DTEND:19970127T160000
UID:2009-11-17T20:55:59-08:00_428606140@limen.stat.ucla.edu
DESCRIPTION:A major challenge in the analysis of observational studies is t
 he assessment of the likely influence of unmeasured potential confounding f
 actors on the estimated effect of the principal causal factor of interest. 
 Two leading classes of models for such data are selection models (developed
  by the economist J Heckman) and counterfactual models (dating back to J Ne
 yman\, more recently developed by D Rubin). The fully Bayesian analysis of 
 such models with Markov Chain Monte Carlo methods poses several technical p
 roblems\, including multi-modality of the posterior distribution and extrem
 ely high serial correlation of the Monte Carlo draws\, and informative prio
 r distributions must be elicited for parameters not well addressed by the d
 ata. In this talk I will illustrate the solution of these problems with exa
 mples from medicine and psychobiology.
SUMMARY:Causal Inference Via Markov Chain Monte Carlo
DTSTART:19970127T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:387
SEQUENCE:0
DTEND:19961104T160000
UID:2009-11-17T20:55:59-08:00_441506403@limen.stat.ucla.edu
DESCRIPTION:Time pressure usually demands that Statisticians analyze data s
 ets quickly\, resulting in the use of models and methods that are simple an
 d easily available in standard software.  Occasionally we have time to take
  one data set thoroughly apart so as to make an attempt at getting the mode
 l completely correct.  This talk sketches a portion of a long term analysis
  of a Pediatric pain repeated measures data set.  The data consists of up t
 o 4 repeated pain tolerance trials on 64 kids.  Pain tolerance is assessed 
 by the length of time the children can keep an arm immersed in cold water. 
  Some of the problems with the data set are (a) some missing data\, (b) cen
 soring of large responses\, (c) repeated measures\, (d) non-constant varian
 ce\, (e) not overwhelming statistical significance and (f) interest in desi
 gning a new study.  Time won't permit the discussion of all of these proble
 ms\, but we will discuss the path that my data analyses have taken and migh
 t take\, and some of the Bayesian methods developed to handle the various p
 roblems.
SUMMARY:Getting it Right? Leisurely Analysis of a Repeated Measures Data Se
 t
DTSTART:19961104T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:388
SEQUENCE:0
DTEND:19961028T160000
UID:2009-11-17T20:55:59-08:00_661347956@limen.stat.ucla.edu
DESCRIPTION:We consider the problem of asymptotically minimax estimation of
  an unknown regression function which belongs to a Holder function class. V
 arious risk measures are discussed\, special attention is given to the Baha
 dur minimax risk. We study kernel estimators expressed via the solution of 
 a certain deterministic optimization problem introduced to the nonparametri
 cs by D.Donoho. The properties of this solution are discussed.
SUMMARY:Minimax Regression Estimation for Holder Function Classes
DTSTART:19961028T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:389
SEQUENCE:0
DTEND:19961021T160000
UID:2009-11-17T20:55:59-08:00_372068515@limen.stat.ucla.edu
DESCRIPTION:This talk covers some of the computer graphical tools we develo
 ped recently for visualizing multivariate data sets. \n\n The first tool is
  a new environment for MDS (multidimensional scaling) model fitting and dia
 gnostics. A dynamic and colored stimulus plot is employed to link the input
  measurements to the output configuration. Classical diagnostic plots (line
 ar\, nonlinear\, transformation) are also equipped with similar linkage fun
 ction for better diagnostics. \n\n The second tool\, GAP (generalized assoc
 iation plots)\, utilizes the convergent property of iteratively formed corr
 elation matrices to study information embedded in a multivariate data set. 
 Various static and dynamic computer graphics are created to retrieve differ
 ent pieces of information embedded in the raw data matrix and proximity mat
 rices for variables and subjects. \n\n Possible extensions of these tools f
 or studying longitudinal multivariate data set will also be discussed. A da
 ta set from the "Taiwan Multidimensional psychopathological group research 
 program (MPGRP)" will be used to illustrate the concept of these new graphi
 cal environment. \n\n Please refer to <a href="http://www.stat.ucla.edu/~ch
 unhouh">Chun-houh Chen's</a> homepage for graphics and additional informati
 on.
SUMMARY:Some Dynamic Graphics for Visualizing Multivariate Data Set
DTSTART:19961021T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:390
SEQUENCE:0
DTEND:19961014T160000
UID:2009-11-17T20:55:59-08:00_541256877@limen.stat.ucla.edu
DESCRIPTION:Q: Are Bayes rules admissible? \n\n A: Almost. \n\n It is known
  that if a decision rule is Bayes against a prior Pi\, then it must also be
  almost-Pi-admissible. We start by presenting a theorem which generalizes t
 his fact and connects it to the stepwise Bayes procedure\, which has been u
 sed to prove admissibility in finite population sampling. We then use this 
 result to prove admissibility of some deision rules in finite population sa
 mpling\, where we consider two problems: \n\n * distributional inference un
 der a strictly proper scoring rule\; \n\n * estimating the population mean 
 under square error loss by sequential sampling.
SUMMARY:Some Admissible Decision Rules in Finite Population Sampling
DTSTART:19961014T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:392
SEQUENCE:0
DTEND:19960930T160000
UID:2009-11-17T20:55:59-08:00_804891844@limen.stat.ucla.edu
DESCRIPTION:We consider linear calibration in the context of repeated measu
 res.  Suppose two scalar quantities x and Y are related by Y = a + b.x + er
 ror\, and that at the calibration step repeated measurements on both x and 
 Y are available for a number of sampling units.  At the prediction step a n
 ewY measurement (possibly together with previous measurements for the new u
 nit) is available\, and it is desired to estimate the corresponding unknown
  x value. \n\n Examples are x = week of pregnancy and Y = fetal ultrasound 
 bone measurement for a number of women during pregnancy monitoring\; and tw
 o indices of renal function (x precise yet expensive\, Y inexact yet cheape
 r) for cancer patients at different times during the course of chemotherapy
 . \n\n We allow for both the intercept a and the slope b to vary between un
 its\, resulting in a random regression coefficient model at the calibration
  step. This complicates the prediction step\, since the unknown x affects t
 he covariance structure of Y.  We discuss point estimation and obtain a  co
 nfidence region for x similar to Fieller's.  Our methods are illustrated on
  a set of ultrasound bladder-size measurements.
SUMMARY:Calibration with Repeated Measures
DTSTART:19960930T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:393
SEQUENCE:0
DTEND:19960910T160000
UID:2009-11-17T20:55:59-08:00_612294756@limen.stat.ucla.edu
DESCRIPTION:We have determined the make-up and impact of several industrial
  sources of organic gases in Houston\, Texas from ambient measurements and 
 found these to be largely incompatible with self-reported emissions of orga
 nic gases in the area.  These gases are man-made precursors to photochemica
 l smog formation. \n\n Regulatory agencies and photochemical models rely on
  these self-reported industrial emission rates\, which are often outdated\,
  incomplete\, or inaccurate.  Our results provide an independent\, objectiv
 e estimate of industrial source compositions and contributions to organic g
 as pollution. \n\n Our approach uses measurements from an automated gas chr
 omatography monitor at a site near the Houston Ship Channel (a large petroc
 hemical complex) from June to November\, 1993.  Multivariate receptor model
 ing (1) was applied to hourly observations of total non-methane organic car
 bon (TNMOC) and fifty two hydrocarbon compounds from C2 to C9\, as listed i
 n Table 1.  Wind direction analysis of the impact of three sources determin
 ed by multivariate receptor modeling showed that these sources should lie c
 lose to the monitoring site. \n\n However\, we were only able to match the 
 smallest of these sources to a specific industrial facility.  The two large
 st sources were incompatible in direction and composition with emissions re
 ported by industries in the vicinity of the monitoring site.
SUMMARY:Reported Emissions of Organic Gases Are Not Consistent With Observa
 tion
DTSTART:19960910T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:394
SEQUENCE:0
DTEND:19960513T160000
UID:2009-11-17T20:55:59-08:00_192979408@limen.stat.ucla.edu
DESCRIPTION:Reconciling competing interests for data is one of the great ch
 allenges in designing a research program.  This talk will present two examp
 les of this statistical balancing act\, one in sampling and estimation\, th
 e other in forecasting. \n\n The sampling problem is a classic one: two dif
 ferent groups want to conduct a survey of a population\, but they have conf
 licting goals.  User group one wants to estimate a set of totals from the p
 opulation\, while user group two cares nothing about the population as a wh
 ole but wants to target subpopulations.  Both groups want the survey to be 
 inexpensive and quick\, so it is possible to get them to agree with each ot
 her (but not on a topic that might help you as the designer).  An approach 
 to the problem using traditional and adaptive sampling is discussed. \n\n T
 he forecasting problem is similar in concept: our desire is to predict bank
  failures.  But do we want to predict which specific banks will fail in the
  next year or do we want to predict the total loss associated with all bank
  failures in a year?  The answer\, of course\, is both!  The problem is how
  to develop a consistent model that concerns itself with both objective fun
 ctions.
SUMMARY:Reconciling Competing Interests in Research Designs
DTSTART:19960513T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:395
SEQUENCE:0
DTEND:19960509T160000
UID:2009-11-17T20:55:59-08:00_665920827@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Teaching Elementary Statistics at Berkeley
DTSTART:19960509T150000
DTSTAMP:20091117T205559
LOCATION:8145 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:396
SEQUENCE:0
DTEND:19960518T160000
UID:2009-11-17T20:55:59-08:00_186468014@limen.stat.ucla.edu
DESCRIPTION:Information-theoretic tools are used to derive minimax risk bou
 nds for density estimation. A metric entropy condition alone determines the
  minimax rate of convergence in each class of density functions. To achieve
  the minimax rates simultaneously for multiple function classes\, we consid
 er  lists of finite-dimensional approximating models and  use model selecti
 on criteria to select adaptively a good model based on data. The use of man
 y candidate models\, as in the case of subset selection\, provides more fle
 xibility for adaptation\, yet significant selection bias can occur with cri
 teria such as AIC. We incorporate a model complexity term in the model sele
 ction criteria to handle this selection bias. It is shown that the risk of 
 the estimated density is bounded by an index of resolvability\, which chara
 cterizes the best tradeoff among approximation error\, estimation error\, a
 nd model complexity. As an application\, we show that the optimal rate of c
 onvergence is simultaneously achieved for density in the Sobolev space with
 out knowing the smoothness parameter <b>s</b> and norm parameter <b>U</b> i
 n advance.
SUMMARY:Model Selection for Density Estimation
DTSTART:19960518T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:401
SEQUENCE:0
DTEND:19960129T160000
UID:2009-11-17T20:55:59-08:00_946575145@limen.stat.ucla.edu
DESCRIPTION:Direct generalization of M-theorem in finite dimensional spaces
  (Pakes and Pollard(1989)) to functional M-theorem in infinite dimensional 
 spaces (Van der Vaart(1995)) may lead to a dilemma even for some smooth mod
 els linearly parametrized: the Fréchet differentiability of likelihood equa
 tion and the invertibility of the derivative operator can not be establishe
 d with respect to the same norm. Since the differentiability and the invert
 ibility are both sufficient conditions in the M-theorem\, this dilemma rend
 ers the M-theorem unverifiable without imposing further conditions. \n\n In
  this paper we present a functional M-theorem and a bootstrap limit theorem
  for linearly parametrized models. For this class of models\, an identity d
 erived from the likelihood equation provides a direct way of proving asympt
 otics. The linearity eliminates the gap between consistency and an <span cl
 ass='math'>O_{P}(n^{-1/2})</span> rate of convergence and asymptotic normal
 ity can be obtained via the continuity of the inverse Fréchet derivative. A
  feature of this type of limit theorems is that the asymptotic normality or
  the validity of nonparametric bootstrap can be established with respect to
  a possibly different norm from the one with respect to which the Fréchet d
 ifferentiability has to be established. A key result worth mentioning that 
 transfers an M-theorem to its bootstrap limit theorem is that the usual sto
 chastic equicontinuity implies the bootstrap equicontinuity under a mild in
 tegrability condition for a natural envelope function. This result makes th
 e bootstrap limit theorem almost a preservation theorem of the functional M
 -theorem we obtain. Two examples are included for demonstration.
SUMMARY:Exploring Linearity in Functional M-Estimation
DTSTART:19960129T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:398
SEQUENCE:0
DTEND:19960212T160000
UID:2009-11-17T20:55:59-08:00_822439337@limen.stat.ucla.edu
DESCRIPTION:Hot deck imputation for nonrespondents is often used in surveys
 . It is a common practice to treat the imputed values as if they are true v
 alues\, and compute survey estimators and their variance estimators using s
 tandard formulas. The variance estimators\, however\, have seriously negati
 ve biases when the rate of nonresponse is appreciable. Methods such as the 
 multiple imputation and the adjusted jackknife have been proposed to obtain
  improved variance estimators. However\, the multiple imputation requires t
 hat multiple data sets be generated and maintained and that the imputation 
 procedure be proper\; the adjusted jackknife requires identification flags 
 to identify nonrespondents. In many practical problems there is only a sing
 le imputed data set in which the nonrespondents are not identifiable (no id
 entification flag). In this paper we derive some asymptotically design-cons
 istent inference procedures in the situation where a stratified multistage 
 sampling design is used to collect survey data\; the hot deck imputation is
  applied to form a single imputed data set\; the nonrespondents are not ide
 ntifiable\; and the survey estimators under consideration are functions of 
 sample means\, sample quantiles\, and sample low income proportions.
SUMMARY:Inference Based on Complex Survey Data with Nonidentifiable Nonresp
 ondents Imputed by Hot Deck
DTSTART:19960212T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:399
SEQUENCE:0
DTEND:19960208T160000
UID:2009-11-17T20:55:59-08:00_97920284@limen.stat.ucla.edu
DESCRIPTION:We will present some new results on Monte Carlo simulation and 
 global optimization. We will follow the traditional approach of constructin
 g a Markov chain with a desired stationary distribution. However\, the conc
 epts of weighting and control will be introduced into the Markov framework.
  We will also discuss the use of "sequential build-up" techniques in Monte 
 Carlo. The methods will be illustrated using the Traveling Salesman Problem
 .
SUMMARY:Markov Chain Monte Carlo and the Salesman Problem
DTSTART:19960208T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:400
SEQUENCE:0
DTEND:19960205T160000
UID:2009-11-17T20:55:59-08:00_789238931@limen.stat.ucla.edu
DESCRIPTION:The satellite-based instrument TOMS (TOTAL OZONE MAPPING SPECTR
 OMETER) measures total column ozone with high precision and high spatial an
 d temporal resolutions. We will view the total column ozone field as a rand
 om field\, and develop physically based space-time correlation models of th
 e total column ozone field that fit the data well and whose parameter value
 s have physical interpretations.
SUMMARY:Modeling the Correlation Structure of the TOMS Ozone Data
DTSTART:19960205T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:402
SEQUENCE:0
DTEND:19960122T160000
UID:2009-11-17T20:55:59-08:00_663331614@limen.stat.ucla.edu
DESCRIPTION:As part of a large study of early pregnancy loss\, urinary prog
 esterone profiles are measured over the course of several menstrual cycles 
 for patients from an artificial insemination clinic. The progesterone measu
 rements from one cycle form a curve\, and together the cycles form a nested
  sample of curves: cycles nested within women who in turn may be nested wit
 hin groups. The urinary profiles are noisy indicators of hormone activity\,
  and missing data\, varying cycle lengths\, and differing numbers of cycles
  per woman also complicate the data. A correlated smoothing spline model ap
 propriate for smoothing this and other nested samples of curves is introduc
 ed. Theoretical and computational issues are discussed\, and performance is
  studied with applications to the progesterone data.
SUMMARY:Modeling a Nested Sample of Curves with Correlated Smoothing Spline
DTSTART:19960122T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:407
SEQUENCE:0
DTEND:19951211T160000
UID:2009-11-17T20:55:59-08:00_255227966@limen.stat.ucla.edu
DESCRIPTION:Criterion functions that are approximately parametric likelihoo
 ds play an increasingly important role in statistics. This general area inc
 ludes modifications of profile likelihoods (Barndorff-Nielsen\, Cox and Rei
 d\, McCullagh and Tibshirani)\, moment based criterion functions (quasi- an
 d projective likelihood) and what one might call almost-likelihoods (sieve 
 likelihood\, bootstrap likelihood). Via 'dual' criterion functions\, certai
 n nonparametric (empirical\, point process) likelihoods can also be conside
 red as part of this family. \n\n One of the questions raised by this develo
 pment is to what extent the excellent small sample accuracy properties of t
 he likelihood ratio statistic and its signed square root (R) are retained f
 or such approximate likelihoods. As a way of addressing this question\, we 
 discuss what asymptotic properties are special to R\, in the hope that such
  properties may have predictive power for small samples. \n\n It turns out 
 that the cumulants of R &mdash\; all of them &mdash\; decay faster than wha
 t can be expected from an asymptotically normal statistic. This relates bot
 h to large and small deviation properties. Affine (Bartlett) correctability
  is a special case corresponding to the first four cumulants. \n\n For appr
 oximate likelihoods\, the number of cumulants displaying this behavior can 
 be characterized by how many Bartlett identities the 'log likelihood' satis
 fies to first order. It turns out\, for example\, that the R-statistic in e
 mpirical likelihood\, though affinely correctable\, does not in general beh
 ave like the R-statistic of a true likelihood. \n\n As an experimental way 
 of generating R-statistics with high accuracy in general inference situatio
 ns\, we look at a class of criterion functions based on resampling methods 
 ('design-your-own' likelihoods).
SUMMARY:Approximate and Artificial Likelihoods
DTSTART:19951211T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:404
SEQUENCE:0
DTEND:19960119T170000
UID:2009-11-17T20:55:59-08:00_461199465@limen.stat.ucla.edu
DESCRIPTION:Inverse problems have been well studied by numerical analysts. 
 In this talk\, we will show that many classical problems in statistics are 
 in fact ill-posed inverse problems. Some methodology elaborated in function
 al analysis together with techniques known from statistics may then be appl
 ied to solve some challenging inverse problems in Statistics.
SUMMARY:Inverse Problems in Statistics
DTSTART:19960119T160000
DTSTAMP:20091117T205559
LOCATION:8145 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:405
SEQUENCE:0
DTEND:19960108T160000
UID:2009-11-17T20:55:59-08:00_322339859@limen.stat.ucla.edu
DESCRIPTION:U.S. Congressmembers and state legislators are elected from dis
 tricts whose boundaries are redrawn every ten years. Sometimes the district
 s lines are drawn by Democratic legislators\, sometimes by Republicans\, an
 d sometimes by ostensibly nonpartisan organizations such as state courts. F
 or two hundred years\, politicians have been accusing each other of "gerrym
 andering": drawing district lines to unfairly favor their party and to make
  safer seats for incumbent politicians. \n\n How effective is gerrymanderin
 g in practice? We analyze 267 elections in thirty U.S. state legislatures s
 ince the 1960s\, including sixty cases of redistricting\, to see how effect
 ive the politicians have been at protecting their own seats and gaining an 
 advantage over the opposition party. This analysis has the form of an obser
 vational study\, in which the treatments are redistricting (under Democrati
 c\, Republican\, or bipartisan control)\, the units are state legislative e
 lections\, and the outcomes are measures of competitiveness and partisan bi
 as that we estimate for each election. \n\n We conclude with a discussion o
 f why fancy statistical modeling (we fit a random effects regression model 
 to each of the 267 statewide elections in our dataset) can help answer subs
 tantive questions in political science. \n\n This work is joint with Gary K
 ing\, Department of Government\, Harvard University.
SUMMARY:Enhancing Democracy Through Legislative Redistricting
DTSTART:19960108T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:406
SEQUENCE:0
DTEND:19951218T160000
UID:2009-11-17T20:55:59-08:00_174726031@limen.stat.ucla.edu
DESCRIPTION:Smoothing parameter selection is among the most intensively stu
 died subjects in nonparametric function estimation. A closely related issue
 \, that of identifying a proper index for the smoothing parameter\, is howe
 ver largely neglected in the existing literature. Through heuristic argumen
 ts and simple simulations\, we shall illustrate that the "default" working 
 indices are often conceptually "incorrect"\, in the sense that they are not
  interpretable across-replicate in repeated experiments\, and as a conseque
 nce\, a few popular working concepts\, such as the expected mean square err
 or and the "degrees of freedom"\, appear vulnerable under close scrutiny. D
 ue to technical constraint\, the arguments are mainly developed in the pena
 lized likelihood setting\, but parallels can be drawn to other settings as 
 well. The development stems from an attempt to understand the well-publiciz
 ed negative correlation between optimal and cross-validation smoothing para
 meters\, which however turns out to bear little statistical relevance.
SUMMARY:Model Indexing and Smoothing Parameter Selection in Nonparametric F
 unction Estimation
DTSTART:19951218T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:411
SEQUENCE:0
DTEND:19951027T160000
UID:2009-11-17T20:55:59-08:00_291939930@limen.stat.ucla.edu
DESCRIPTION:In likelihood ratio tests involving inequality-constrained hypo
 theses\, (the multivariate analog of one-sided tests) the standard Neyman-P
 earson test guarding against all parameter points in a compound null hypoth
 esis can be extremely conservative. The ordinary parametric bootstrap is ge
 nerally inconsistent and usually too liberal. Two methods of correcting the
  inconsistency of the parametric bootstrap are proposed: shrinking the cons
 traint set toward the maximum likelihood estimate and superefficient estima
 tion of the active set of constraints. In shrinkage adjustment\, the shrink
 age can be chosen optimally using bootstrap calibration. These methods are 
 compared with the double bootstrap\, the subsampling bootstrap\, Bayes fact
 ors\, and Bayesian P-values. The double bootstrap seems to work\, despite b
 eing inconsistent. The Bayesian methods are extremely liberal.
SUMMARY:Inequality Constraints and Significance
DTSTART:19951027T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:409
SEQUENCE:0
DTEND:19951120T160000
UID:2009-11-17T20:55:59-08:00_422640196@limen.stat.ucla.edu
DESCRIPTION:We consider linear approximation (LA) confidence intervals for 
 functions of the parameters in a nonlinear regression model. These interval
 s are extensively used\, but at times have coverage probabilities that diff
 er significantly from there nominal values. We give a diagnostic plot and i
 ndex to detect these failures and show how to use the diagnostics to estima
 te actual coverage probabilities and suggest transformations that will impr
 ove the LA intervals. The coverage probability estimates provide a natural 
 calibration for the diagnostics. These results are motivated by theorems. T
 heir performance under less than ideal conditions is demonstrated by a simu
 lation study using a variety of nonlinear regression problems. Additionally
  we show how to calibrate other diagnostics and how to use them to suggest 
 transformations. These include the profile t plot and asymmetry and bias in
 dices.
SUMMARY:Diagnostics in Transformations for Linearization Confidence Interva
 ls in Nonlinear Regression
DTSTART:19951120T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:410
SEQUENCE:0
DTEND:19951030T160000
UID:2009-11-17T20:55:59-08:00_728051041@limen.stat.ucla.edu
DESCRIPTION:Additive regression models have turned out to be a useful stati
 stical tool in the analysis of high dimensional data sets. Recently\, an es
 timator of additive components has been introduced by Linton and Nielsen (1
 994) which is based on marginal integration. The explicit definition of thi
 s estimator makes possible a fast computation and allows an asymptotic dist
 ribution theory. In this talk a modification of this procedure is introduce
 d. We propose to use local linear fits instead of kernel smoothing and to i
 ntroduce a weight function. We demonstrate that with an appropriate choice 
 of the weight function\, the additive components can be efficiently estimat
 ed---each additive component can be estimated as well as if the rest of com
 ponents were known. Application of local linear fits reduces the design rel
 ated bias. Estimaiton of parametric components as well as nonparametric com
 ponents in the additive partially linear models will also be addressed.
SUMMARY:Direct Estimation of Additive and Linear Components for High Dimens
 ional Data
DTSTART:19951030T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:412
SEQUENCE:0
DTEND:19951016T160000
UID:2009-11-17T20:55:59-08:00_125380881@limen.stat.ucla.edu
DESCRIPTION:It is known that certain combinations of one-sided sequential p
 robability ratio tests are asymptotically optimal (relative to expected sam
 ple size) for problems involving a finite number of possible distributions 
 when probabilities of errors tend to zero and observations are independent 
 and identically distributed according to one of the underlying distribution
 s. The objective of this talk is to show that two specific constructions of
  sequential tests asymptotically minimize not only expected time of observa
 tion but also any positive moment of observation time under fairly general 
 conditions for finite number of simple hypotheses. This result appears to b
 e true for general statistical models which include correlated and nonhomog
 eneous processes observed either in discrete or continuous time. For the pr
 oblems with nuisance parameters\, we consider invariant sequential tests an
 d show that the same result is valid for this case. Finally\, we apply gene
 ral results to the solution of several particular problems such as multi-st
 ate slippage problem for correlated Gaussian processes and for statistical 
 models with nuisance parameters.
SUMMARY:Asymptotic Optimality of Certain Multi-Alternative Sequential Tests
  for Correlated and Non-stationary Processes
DTSTART:19951016T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:413
SEQUENCE:0
DTEND:19951002T160000
UID:2009-11-17T20:55:59-08:00_683185294@limen.stat.ucla.edu
DESCRIPTION:NA
SUMMARY:Smoking Causes Cancer
DTSTART:19951002T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:3
SEQUENCE:0
DTEND:20090519T160000
UID:2009-11-17T20:55:59-08:00_521239278@limen.stat.ucla.edu
DESCRIPTION:The subject of stochastic approximation was founded by Robins a
 nd Monro (1951). After five decades of continual development\, it has devel
 oped into an important area in systems control and optimization\, and it ha
 s also served as a prototype for the development of adaptive algorithms for
  on-line estimation and control of stochastic systems. Recently\, it has be
 en used in statistics with Markov chain Monte Carlo for solving maximum lik
 elihood estimation problems and for general simulation and optimizations. I
 n this talk\, we first show that the trajectory averaging estimator is asym
 ptotically efficient for the stochastic approximation MCMC (SAMCMC) algorit
 hm under mild conditions\, and then apply this result to the stochastic app
 roximation Monte Carlo algorithm (Liang et al.\, 2007). The theoretical res
 ult is illustrated by a numerical example\, which indicates that for the st
 ochastic approximation Monte Carlo algorithm\, the trajectory averaging est
 imator can be generally superior to the conventional estimator in terms of 
 bias and variance. The application of the trajectory averaging estimator to
  other stochastic approximation MCMC algorithms\, e.g.\, a stochastic appro
 ximation MLE algorithm for missing data problems\, will also be considered 
 in the paper.
SUMMARY:Trajectory Averaging for Stochastic Approximation MCMC Algorithms
DTSTART:20090519T150000
DTSTAMP:20091117T205559
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:5
SEQUENCE:0
DTEND:20090505T160000
UID:2009-11-17T20:55:59-08:00_228905822@limen.stat.ucla.edu
DESCRIPTION:In this talk\, we introduce and study multivariate spacings. Th
 e spacings are developed using the order statistics derived from data depth
 . Specifically\, the spacing between two consecutive order statistics is th
 e region which bridges the two order statistics\, in the sense that the reg
 ion contains all the points whose depth values fall between the depth value
 s of the two consecutive order statistics. These multivariate spacings can 
 be viewed as a data-driven realization of the so-called "statistically equi
 valent blocks". These spacings assume a form of center-outward layers of "s
 hells" ("rings" in the two-dimensional case)\, where the shapes of the shel
 ls follow closely the underlying probabilistic geometry. The properties and
  applications of these spacings are studied. In particular\, the spacings a
 re used to construct tolerance regions. The construction of tolerance regio
 ns is nonparametric and completely data driven\, and the resulting toleranc
 e region reflects the true geometry of the underlying distribution. This is
  different from most existing approaches which require that the shape of th
 e tolerance region be specified in advance. The proposed tolerance regions 
 are shown to meet the prescribed specifications\, in terms of β-content and
  β-expectation. They are also asymptotically minimal under elliptical distr
 ibutions. Finally\, we present a simulation and comparison study on the pro
 posed tolerance regions. \n\n This is joint work with Prof. Regina Y. Liu f
 rom Rutgers University.
SUMMARY:Multivariate Spacings Based on Data Depth and Construction of Nonpa
 rametric Multivariate Tolerance Regions
DTSTART:20090505T150000
DTSTAMP:20091117T205559
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:8
SEQUENCE:0
DTEND:20090414T160000
UID:2009-11-17T20:55:59-08:00_597171636@limen.stat.ucla.edu
DESCRIPTION:A fundamental problem in biology is to understand how the cell 
 regulates its genes. Gene transcription through the interaction between tra
 nscription factors and their binding sites is a crucial step in the regulat
 ion process. The pattern of transcription factor binding sites (TFBS's) is 
 named as a motif\, and the procedure for detecting TFBS's is referred as to
  motif finding\, which is an important problem in computational biology. Mo
 tif finding can be formulated as detecting signal from background by treati
 ng a TFBS as signal and its surrounding DNA sequences as background. Due to
  heterogeneous (nucleotide) base composition across different sequence regi
 ons\, the common assumption of homogeneous background deserves serious chec
 king for its effects on motif finding. The heterogeneity in background beco
 mes even more complicated when evolutionary conservation in multiple alignm
 ents is utilized to help motif finding. In this study\, we propose a genera
 tive model framework to capture the heterogeneity of base composition and e
 volutionary conservation in multiple alignments simultaneously to build pro
 per background for multiple species motif finding. Simulation studies and e
 xperimental results from biological data sets demonstrate that the proposed
  framework can substantially improve motif finding performance.
SUMMARY:Heterogeneity in DNA Multiple Alignments: Modeling\, Inference\, an
 d Applications
DTSTART:20090414T150000
DTSTAMP:20091117T205559
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:15
SEQUENCE:0
DTEND:20090217T160000
UID:2009-11-17T20:55:59-08:00_966085914@limen.stat.ucla.edu
DESCRIPTION:A predictor variable or dose that is measured with substantial 
 error may possess an error-free milestone\, such that it is known with negl
 igible error whether the value of the variable is to the left or right of t
 he milestone. Such a milestone provides a basis for estimating a linear rel
 ationship between the true but unknown value of the error-free predictor an
 d an outcome\, because the milestone creates a strong and valid instrumenta
 l variable. The inferences are nonparametric and robust\, and in the simple
 st cases\, they are exact and distribution free. We also consider multiple 
 milestones for a single predictor and milestones for several predictors who
 se partial slopes are estimated simultaneously.  Examples are drawn from th
 e Wisconsin Longitudinal Study\, in which a BA degree acts as a milestone f
 or sixteen years of education\, and the binary indicator of military servic
 e acts as a milestone for years of service.  This is joint work with Paul R
 osenbaum.
SUMMARY:Error Free Milestones in Error Prone Measurements
DTSTART:20090217T150000
DTSTAMP:20091117T205559
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:16
SEQUENCE:0
DTEND:20090210T160000
UID:2009-11-17T20:55:59-08:00_598943809@limen.stat.ucla.edu
DESCRIPTION:Numerical models are widely used to understand and predict spat
 io-temporal processes.  While the specific aspects of individual models are
  different\, they all share a similar feature. They are deterministic model
 s that mathematically approximate the underlying physical and chemical proc
 esses via nonlinear partial differential equations. Their predictions are g
 iven in terms of averages over grid cells and\, being derived under a deter
 ministic paradigm\, do not convey any information about the inherent uncert
 ainty.  In this talk\, I will present two approaches for stochastic predict
 ion using output from numerical models. \n\n The first is an attractive spa
 tio-temporal model that allows to downscale output from numerical models to
  point level\, thus offering a solution to the problem of spatial misalignm
 ent between observational data and computer model output.  The second is an
  example\, developed specifically for numerical weather forecasts of precip
 itation accumulation\, to postprocess the  output from numerical weather pr
 ediction models and produce probabilistic forecasts of precipitation accumu
 lation at multiple sites simultaneously. \n\n This is joint work with A. Ge
 lfand\, D. Holland\, A. Raftery and T. Gneiting.
SUMMARY:Stochastic Prediction Using Output from Numerical Models
DTSTART:20090210T150000
DTSTAMP:20091117T205559
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:17
SEQUENCE:0
DTEND:20090203T160000
UID:2009-11-17T20:55:59-08:00_303847509@limen.stat.ucla.edu
DESCRIPTION:I will give a brief overview of the main issues underlying the 
 characterization of climate change projections and their uncertainty\, and 
 then focus on a statistical analysis that combines information from an ense
 mble of climate models. The approach is that of a Bayesian hierarchical mod
 el. By considering the performance of different climate models over differe
 nt regions\, parameters quantifying systematic biases and region-specific e
 ffects can be estimated\, and the posterior distribution of regional climat
 e change signals or the posterior predictive distribution of a "new climate
  model" can be offered as representations of the uncertainty in our future 
 climate. A cross-validation exercise is proposed in place of direct validat
 ion of these "predictions" (in the interest of time\, given that these proj
 ections are aimed at mid-to-end of the century horizons!). If time permits 
 I will briefly mention applications of these probabilistic projections in t
 he areas of agricultural  and hydrological impacts studies\, both examples 
 of research areas that are increasingly sensitive to the need of producing 
 probabilistic information. \n\n BIO: I am a Research Scientist at Climate C
 entral\, a new non-profit organization dedicated to the synthesis and commu
 nication of the science and the solutions of climate change. My research fo
 cuses on the analysis and statistical characterization of climate change pr
 ojections and their uncertainty\, as derived from climate models\, especial
 ly at the regional scale. I am increasingly engaged in impact studies\, esp
 ecially in the hydrological and food security sectors. I began this kind of
  work as a Project Scientist at the National Center for Atmospheric Researc
 h (NCAR) in Boulder\, CO\, where I worked until September 2007 and with who
 se scientists I collaborate regularly. I have been a visiting investigator 
 at the Department of Global Ecology\, Carnegie Institution\, Stanford until
  August 2008 and am now a visiting scientist back at NCAR.  I am a contribu
 ting author of the Intergovernmental Panel on Climate Change\, Fourth Asses
 sment Report\, for Chapter 10\, Global Climate Projections\, and Chapter 11
 \,  Regional Climate Projections\, by Working Group I and Chapter 2\, New A
 ssessment Methods and the Characterization of Future Conditions\, by Workin
 g Group II. \n\n Education:  My undergraduate degree is in Economics with e
 mphasis in Statistics from Universita' L. Bocconi in Milan\, Italy. I recei
 ved a Ph.D. in Statistics and Decision Sciences from Duke University. I was
  a post-doc in the Geophysical Statistics Project at NCAR.
SUMMARY:A Bayesian Approach to Probabilistic Projections of Climate Change
DTSTART:20090203T150000
DTSTAMP:20091117T205559
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:20
SEQUENCE:0
DTEND:20090113T160000
UID:2009-11-17T20:55:59-08:00_962072657@limen.stat.ucla.edu
DESCRIPTION:Suppose that we want to simulate from the posterior density for
  the parameter <code>&#952\;</code> given data <code>y</code>\, with prior 
 <code>p(&#952\;)</code> and likelihood <code>f<sub>&#952\;(y)</sub> = h<sub
 >&#952\;(y)</sub>/c<sub>&#952\;</sub></code>\, where the normalizing consta
 nt <code>c<sub>&#952\;</sub></code> is intractable. Thus the posterior dens
 ity \n\n <code>p(&#952\;|y) &#8733\; p(&#952\;)h<sub>&#952\;(y)</sub>/c<sub
 >&#952\;</sub></code> \n\n is not computable. In an ordinary Metropolis-Has
 tings algorithm for drawing samples from the posterior distribution the acc
 eptance probability depends on the "unknown" ratio of normalizing constants
  <code>c<sub>&#952\;</sub>/c<sub>&#952\;'</sub></code>. Most methods to dat
 e have used various approximations to estimate or eliminate such ratios of 
 normalizing constants. In M√∏ller et al. (Biometrika\, 2006\, pages 451-458
 ) we presented the first Metropolis-Hastings algorithm for drawing samples 
 from the posterior distribution without approximation. It is called the <i>
 auxiliary variable method</i>\, since we extend the posterior distribution 
 by introducing a certain auxiliary variable so that the acceptance probabil
 ity can be computed. The auxiliary variable method is a nice application ex
 ample of perfect simulation algorithms\, and it has e.g. been used for Baye
 sian analysis of Gibbs models (Markov random fields and Markov point proces
 ses). Moreover\, the auxiliary variable method has more recently been modif
 ied and extended to more efficient MCMC algorithms.
SUMMARY:An Auxiliary Variable Method for Metropolis-Hastings Algorithms for
  Distributions with Intractable Normalizing Constants
DTSTART:20090113T150000
DTSTAMP:20091117T205559
LOCATION:A25 Haines Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:23
SEQUENCE:0
DTEND:20081118T160000
UID:2009-11-17T20:55:59-08:00_629785334@limen.stat.ucla.edu
DESCRIPTION:At a depth of about 2890 km\, the core-mantle boundary (CMB) se
 parates turbulent flow of liquid metals in the outer core from slowly conve
 cting\, highly viscous mantle silicates. The CMB marks the most dramatic ch
 ange in dynamic processes and material properties in our planet\, and accur
 ate images of the structure at or near the CMB&mdash\;over large areas&mdas
 h\;are crucially important for our understanding of present day geodynamica
 l processes and the thermo-chemical structure and history of the mantle and
  mantle-core system. In addition to mapping the CMB we need to know if othe
 r structures exist directly above or below it\, what they look like\, and w
 hat they mean (in terms of physical and chemical material properties and ge
 odynamical processes). Detection\, imaging\, (multi-scale) characterization
 \, and understanding of structure (e.g.\, interfaces) in this remote region
  have been -- and are likely to remain -- a frontier in cross-disciplinary 
 geophysics research. I will discuss the statistical problems and challenges
  in imaging the CMB through generalized Radon transform.
SUMMARY:A Journey to the Center of the Earth
DTSTART:20081118T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:26
SEQUENCE:0
DTEND:20081014T160000
UID:2009-11-17T20:55:59-08:00_279152370@limen.stat.ucla.edu
DESCRIPTION:Changepoint problems deal with detecting abrupt changes in the 
 state of a process\, where information one has about the state of affairs i
 s in the form of observations. In the sequential setting\, as long as the b
 ehavior of observations is consistent with the "normal" state\, one is cont
 ent to let the process continue. If the state changes\, then one is interes
 ted in detecting that a change is in effect as soon as possible after its o
 ccurrence. I will consider the simple changepoint problem setting\, where o
 bservations are iid pre-change and iid post-change\, with known pre- and po
 st-change distributions. The Shiryaev-Roberts detection procedure is known 
 to be asymptotically minimax in the sense of minimizing maximal expected de
 tection delay subject to a bound on the average run length to false alarm\,
  as the latter goes to infinity (i.e.\, for low false alarm rate). I will p
 resent other optimality properties of the Shiryaev-Roberts procedure. Speci
 fically\, I will first prove that the Shiryaev-Roberts procedure is exactly
  optimal in the sense of minimizing the integral average delay to detection
  for an arbitrary average run length to false alarm. This is instrumental f
 or proving optimality in a more practical setting where a change occurs in 
 a distant future and is preceded by a stationary flow of false alarms. I wi
 ll prove that the Shiryaev-Roberts procedure is the best (exactly) that one
  can do in terms of minimizing the expected detection delay in the latter s
 etting for any false alarm rate. The method of proof relies on optimal stop
 ping theory and on renewal theory.
SUMMARY:Exact Optimality Properties of the Shiryaev-Roberts Procedure for D
 etecting Changes in Distributions
DTSTART:20081014T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:32
SEQUENCE:0
DTEND:20080506T160000
UID:2009-11-17T20:55:59-08:00_196090060@limen.stat.ucla.edu
DESCRIPTION:We propose an active basis model\, a shared sketch algorithm\, 
 and a computational architecture of sum-max maps for representing\, learnin
 g\, and recognizing deformable templates. In our generative model\, a defor
 mable template is in the form of an active basis\, which consists of a smal
 l number of Gabor wavelet elements at selected locations and orientations. 
 These elements are allowed to slightly perturb their locations and orientat
 ions before they are linearly combined to generate the observed image. The 
 active basis model\, in particular\, the locations and the orientations of 
 the basis elements\, can be learned from training images by the shared sket
 ch algorithm. The algorithm selects the elements of the active basis sequen
 tially from a dictionary of Gabor wavelets at a dense collection of locatio
 ns and orientations. When an element is selected at each step\, the element
  is shared by all the training images\, and the element is perturbed to enc
 ode or sketch a nearby edge segment in each training image. The recognition
  of the deformable template from an image can be accomplished by a computat
 ional architecture that alternates the sum maps and the max maps. The compu
 tation of the max maps deforms the active basis to match the image data\, a
 nd the computation of the sum maps scores the template matching by the log-
 likelihood of the deformed active basis. \n\n Joint work with Ying Nian Wu\
 , Haifeng Gong\, and Song-Chun Zhu.
SUMMARY:Active Basis for Modeling\, Learning and Recognizing Deformable Tem
 plates
DTSTART:20080506T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:34
SEQUENCE:0
DTEND:20080422T160000
UID:2009-11-17T20:55:59-08:00_423185933@limen.stat.ucla.edu
DESCRIPTION:I will discuss methods for estimation and interpolation of para
 meters of geophysics-based statistical models for predicting nonstationary 
 and nonhomogeneous space-time marked point process. Spline functions are us
 ed to characterize the evolution and variation of the parameter in time and
  space\, respectively. Since many coefficients of the spline functions are 
 required\, I use the penalized log-likelihood with the standard roughness p
 enalties for the spline functions to obtain sensible estimates. The penaliz
 ed log-likelihood is interpreted by the Bayesian framework\, and weights fo
 r the penalties are adjusted objectively by maximizing the integrated poste
 rior function. Comparison of priors includes isotropic versus anisotropic t
 he roughness penalties. The current methods and models are recently applied
  to the early forecasting of aftershock probability where the data are only
  partially available immediately after the main shock. \n\n CONTENT: Bayesi
 an Analysis of biases\, Nonstationary and anisotropic statistical models wi
 th the nonuniform detection rates in time and space\, Early forecasting of 
 aftershock probability.
SUMMARY:Modeling of Heterogeneous Datasets
DTSTART:20080422T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:37
SEQUENCE:0
DTEND:20080401T160000
UID:2009-11-17T20:55:59-08:00_728152115@limen.stat.ucla.edu
DESCRIPTION:Environmental datasets obtained from satellites are typically m
 assive in size. The massiveness causes problems in computing optimal spatia
 l (kriging) predictors. In this talk\, a flexible family of nonstationary c
 ovariance functions is constructed using a set of basis functions that is f
 ixed in number. This results in computational simplications in deriving the
  kriging predictor and its kriging variance. We call the methodology fixed 
 rank kriging (FRK) and we apply it to a large dataset of remotely sensed To
 tal Column Ozone (TCO) data\, observed over the entire globe. This talk rep
 resents joint research with Gardar Johannesson. \n\n Dr. Cressie is Directo
 r of the Program in Spatial Statistics and Environmental Sciences\, Ohio St
 ate University.
SUMMARY:Spatial Prediction for Massive Datasets
DTSTART:20080401T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:39
SEQUENCE:0
DTEND:20080226T160000
UID:2009-11-17T20:55:59-08:00_367920562@limen.stat.ucla.edu
DESCRIPTION:The sequential Monte Calo (SMC) methodologies have been shown t
 o have great promises in solving very high dimensional and complex problems
  often encountered in applications such as communication\, bioinformatics a
 nd financial data analysis. The key to a successful SMC implementation is e
 fficiency\, not only in terms of statistical inference accuracy\, but also 
 on the computational complexity. Efficiency is directly related to the desi
 gn of the key components of SMC\, including the intermediate distributions\
 , the trial 'growth' distribution\, and the resampling method. \n\n Many pr
 oblems in application share a common feature - the target distribution is h
 ighly constrained. That is\, the target distribution is a truncated distrib
 ution on an ill-shaped subspace of a high dimensional space. The constraint
 s\, without careful treatments\, are a main source of obstacles in successf
 ul implementations of SMC. In this talk\, we develop a set of algorithms ca
 tegorized as Constrained Sequential Monte Carlo (CSMC) for solving such pro
 blems\, including strategies in designing the intermediate distributions\, 
 the trial distributions\, the resampling steps and Markov moves with CSMC.
SUMMARY:Constrained Sequential Monte Carlo (CSMC)
DTSTART:20080226T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:45
SEQUENCE:0
DTEND:20080115T160000
UID:2009-11-17T20:55:59-08:00_26191067@limen.stat.ucla.edu
DESCRIPTION:The notion of using context information for solving high-level 
 vision problems has been increasinly realized in the field. However\, how t
 o learn an effective and efficient context model\, together with the image 
 appearance\, remains mostly unknown. The current literature using Markov Ra
 ndom Fields (MRFs) and Conditional Random Fields (CRFs) often invovles spec
 ific algorithm design\, in which the modeling and computing stages are stud
 ied in isolation. In this talk\, I present an auto-context algorithm. Auto-
 context learns an integrated low-level and context model\, and is very gene
 ral and easy to implement. Under nearly the identical parameter setting in 
 the training\, we apply the algorithm on three challenging vision applicati
 ons: horse segmentation\, human body configuration\, and scene region label
 ing. The proposed algorithm outperforms many existing algorithms\, in both 
 speed and quality. Moreover\, the scope of the proposed algorithm goes beyo
 nd high-level vision. It has the potential to be used for a wide variety of
  problems of multi-variate labeling.
SUMMARY:Auto-context and Its Application for High-level Vision Tasks
DTSTART:20080115T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:49
SEQUENCE:0
DTEND:20071120T160000
UID:2009-11-17T20:55:59-08:00_423330323@limen.stat.ucla.edu
DESCRIPTION:I give a finite-sample analysis of predictive inference procedu
 res after model selection in a statistically challenging setting with a pot
 entially infinite number of explanatory variables\, where no regularity con
 ditions are imposed on unknown parameters\, where the number of explanatory
  variables in a "good" model can be of the same order as sample size\, and 
 where the number of candidate models can be of larger order than sample siz
 e.  The performance of inference procedures is evaluated conditional on the
  training sample.  Under weak conditions on only the number of candidate mo
 dels and on their complexity\, and uniformly over all data-generating proce
 sses under consideration\, I show that a certain prediction interval is app
 roximately valid and short with high probability in finite samples\, in the
  sense that its actual coverage probability is close to the nominal one\, a
 nd in the sense that its length is close to the length of an infeasible int
 erval that is constructed by actually knowing the "best" model.  Similar re
 sults are shown to hold for predictive inference procedures other than pred
 iction intervals like\, e.g.\, tests of whether a future response will lie 
 above or below a given threshold. \n\n Bio-sketch: Dr. Leeb received his Ph
 .D in mathematics from University of Salzburg in 1997. From 1998-2002\, he 
 was with the Department of Statistics\, University of Vienna. He joined Yal
 e University in 2002 where he is now an associate professor.
SUMMARY:Conditional Predictive Inference Post Model Selection
DTSTART:20071120T150000
DTSTAMP:20091117T205559
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:54
SEQUENCE:0
DTEND:20071009T160000
UID:2009-11-17T20:55:59-08:00_104958538@limen.stat.ucla.edu
DESCRIPTION:We describe a novel noisy-logical distribution for representing
  the  distribution of a binary output variable conditioned on multiple bina
 ry input variables. The distribution is represented in terms of noisy-or's 
 and noisy-and-not's of  causal features which are conjunctions of the binar
 y inputs. The standard noisy-or and noisy-and-not models\, used in causal r
 easoning and artificial intelligence\, are special cases of the noisy-logic
 al distribution. We prove that the noisy-logical distribution is complete i
 n the sense that it can represent all conditional distributions provided a 
 sufficient number of causal factors are used. We illustrate the noisy-logic
 al distribution by showing that it can account for new experimental finding
 s on how humans perform causal reasoning in more complex contexts. Finally\
 , we speculate on the use of the noisy-logical distribution for causal reas
 oning and artificial intelligence.
SUMMARY:The Noisy-Logical Distribution and its Application to Causal Infere
 nce
DTSTART:20071009T150000
DTSTAMP:20091117T205559
LOCATION:1240B Kinsey Science Teaching Pavilion
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:87
SEQUENCE:0
DTEND:20060523T160000
UID:2009-11-17T20:55:59-08:00_414117125@limen.stat.ucla.edu
DESCRIPTION:Traditional geometric feature inspection or form calibration re
 lies on Coordinate Measuring Machines (CMM) or interferometers. However\, f
 or aspheres and free-form optics emerging in modern technologies\, none of 
 the existing methods works ideally. Recently\, the Manufacturing Engineerin
 g Laboratory of NIST has developed a prototype semi-robotic measuring machi
 ne\, called the Geometry Measuring Machine (GEMM)\, by using the Tyman-Gree
 n phase-measuring interferometer as curvature sensor. In this talk\, I pres
 ent some statistical research results from this multidisciplinary collabora
 tion on the development of metrology for this cutting-edge nanoscale measur
 ement tool.  The statistical issues involve curvature extraction from local
  topographic data from each of the measured grid points on the surface\, an
 d developing an uncertainty theory of profile reconstruction using ordinary
  differential equations based on the definition of curvature for smooth sur
 faces using differential geometry theory. We propose a nonparametric local 
 polynomial regression method for curvature estimation and demonstrate its a
 dvantages over the current approach via the straightforward circle fitting 
 method. If time permits\, I will discuss some open issues including the 3-d
  challenges and a statistical design problem of deciding how many and how t
 o place the sampling points on the testing surface. \n\n (Based on the join
 t work with Dr. Nadia Machkour-Deshayes\, a post-doctoral researcher at NIS
 T\, and other members on the GEMM project with the Manufacturing Engineerin
 g Laboratory of NIST.)
SUMMARY:Statistical Theory for 3-D Topography Using Geometry Measuring Mach
 ines
DTSTART:20060523T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:91
SEQUENCE:0
DTEND:20060502T160000
UID:2009-11-17T20:55:59-08:00_539393826@limen.stat.ucla.edu
DESCRIPTION:To pull a dataset together for a graduate statistics class in 2
 005 I read the tracks from several of my CDs into a music editing program. 
 I snipped out the first 40 seconds of each track\, and converted this to nu
 meric data using the R package\, tuneR. I calculated descriptive statistics
  for each track\, yielding 72 variables for 62 music clips\, allowing us to
  study these questions:</p> \n\n Can I get my computer to recognize Rock fr
 om Classical tracks? \n\n How do New Wave clips compare to Rock and Classic
 al? \n\n What are the similarities between the music from Abba\, Beatles\, 
 Eels\, Vivaldi\, Beethoven\, Mozart\, and Enya? \n\n In this talk we will l
 ook at supervised classification into Rock and Classical classes\, and clus
 ter analysis of the music clips. One particularly interesting part of the c
 luster analysis is to look at self-organizing maps wrapping through high-di
 mensional data. \n\n BIOSKETCH<br/> Professor Cook is a member of the Stati
 stics Department and Virtual Reality Applications Lab at Iowa State Univers
 ity. She is a member of the Human Computer Interaction\, and Bioinformatics
  and Computational Biology Graduate programs. Her research is on methods fo
 r the visualization of high-dimensional data.<br/> PhD from Rutgers Univers
 ity 1993.
SUMMARY:An EDA of My CDs
DTSTART:20060502T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:92
SEQUENCE:0
DTEND:20060425T160000
UID:2009-11-17T20:55:59-08:00_489271733@limen.stat.ucla.edu
DESCRIPTION:The Bidirectional Reflectance Distribution Function (BRDF) desc
 ribes the way a surface reflects light. BRDFs are complex mathematical obje
 cts that\, while allowing for a complete radiometric description of light r
 eflecting from a surface\, can be difficult to use in practice. Recently th
 ere has been interest in understanding the perception of reflectance in a s
 imilar manner to the work done over the last two centuries on the perceptio
 n of color. The aim is to construct a low-dimensional\, perceptual space fo
 r BRDFs that can be easily navigated\, similar to a perceptually uniform co
 lor space.  To this end\, we design and carry out a comprehensive psychophy
 sical study of the perception of measured reflectance. This is the largest 
 study of its kind to date\, and the first to use real material measurements
 . In addition\, we introduce a new multidimensional scaling (MDS) algorithm
  for analyzing ordinal data that unlike existing methods is both efficient 
 and optimal. We use the results of our study to construct a perceptual spac
 e of these BRDFs and introduce a new method for perceptual construction of 
 novel BRDFs. \n\n This is joint work with Josh Wills\, Sameer Agarwal and D
 avid Kriegman. \n\n Biosketch:<br/> Serge Belongie received the B.S. degree
  (with honor) in Electrical Engineering from Caltech in 1995 and the M.S. a
 nd Ph.D. degrees in Electrical Engineering and Computer Sciences (EECS) at 
 U.C. Berkeley in 1997 and 2000\, respectively. While at Berkeley\, his rese
 arch was supported by a National Science Foundation Graduate Research Fello
 wship. He is also a co-founder of Digital Persona\, Inc.\, and the principa
 l architect of the Digital Persona fingerprint recognition algorithm. He is
  currently an assistant professor in the Computer Science and Engineering D
 epartment at U.C. San Diego. His research interests include computer vision
  and pattern recognition. He is a recipient of the NSF CAREER Award and the
  Alfred P. Sloan Research Fellowship. In 2004 MIT Technology Review named h
 im to the list of the 100 top young technology innovators in the world (TR1
 00).
SUMMARY:Toward a Perceptual Space for Reflectance
DTSTART:20060425T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:94
SEQUENCE:0
DTEND:20060411T160000
UID:2009-11-17T20:55:59-08:00_990039145@limen.stat.ucla.edu
DESCRIPTION:In this talk\, I will present a new method for automatic segmen
 tation of heterogeneous image data\, which is very common in medical image 
 analysis. The work takes one step toward bridging two state-of-the-art imag
 e segmentation approaches:  graph-based and generative model-based segmenta
 tion.  Specifically\, the main contribution of the work is a mathematical f
 ormulation for incorporating soft model assignments into the calculation of
  affinities\, which are traditionally model free.  This model-aware affinit
 y measurement has been integrated into the multilevel Segmentation by Weigh
 ted Aggregation algorithm.  As a byproduct of the integrated Bayesian model
  classification\, each region is assigned a most likely model class accordi
 ng to a set of learned model classes.  The technique has been applied to th
 e task of detecting and segmenting brain tumor and edema in multimodal magn
 etic resonance image volumes.  Our results indicate the benefit of incorpor
 ating model-aware affinities into the segmentation process for the difficul
 t case of brain tumor.
SUMMARY:Multilevel Image Segmentation and Integrated Bayesian Model Classif
 ication with an Application to Brain Tumor Imaging
DTSTART:20060411T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:97
SEQUENCE:0
DTEND:20060314T160000
UID:2009-11-17T20:55:59-08:00_611757685@limen.stat.ucla.edu
DESCRIPTION:NASA Earth remote-sensing satellites generate hundreds of gigab
 ytes of data every day\, presenting an enormous challenge for scientists bu
 t a fascinating opportunity for statisticians and computer scientists.  I w
 ill be talking about a couple of different problems I've worked on at JPL t
 hat illustrate some of the interesting data mining challenges: (1) using su
 pport vector machines to classify each observed pixel as cloud\, smoke\, la
 nd\, water\, or ice/snow.  (2) using data fusion and vision algorithms to l
 ook for images of smoke plumes and empirically investigate the relationship
  between fire strength and plume height.  In both of these problems\, many 
 standard statistical and algorithmic techniques have been quite useful\, bu
 t real-world data sets invariably have idiosyncrasies and asymmetries that 
 cause problems for standard techniques.  I will present several examples of
  this\, including many open problems that I am not aware of a good solution
  to yet. \n\n Bio-sketch: Dominic Mazzoni is a senior computer scientist in
  the Machine Learning and Instrument Autonomy Group at the Jet Propulsion L
 aboratory\, California Institute of Technology.  He has a master's degree i
 n computer science from Carnegie Mellon University\, where he worked on mus
 ic information retrieval research.  His research focus is on both pure mach
 ine learning\, and also the application of machine learning algorithms to E
 arth science data.  Outside of JPL\, Mr. Mazzoni is also the founder and le
 ad developer of Audacity\, a popular open-source audio editor.
SUMMARY:Earth Science Data Mining Challenges
DTSTART:20060314T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:99
SEQUENCE:0
DTEND:20060307T160000
UID:2009-11-17T20:55:59-08:00_715162731@limen.stat.ucla.edu
DESCRIPTION:In the first part of the talk I present a novel constructive ap
 proach called HdBCS to generate large-scale undirected Gaussian graphical m
 odels based on a sparse representation of the joint distribution of covaria
 tes via sets of linear regressions. I discuss the validity of my stochastic
  search algorithm and show how to estimate various dependence measures (e.g
 .\, Kendall's tau\, Spearman's rho) by taking into account model uncertaint
 y. I briefly introduce GraphExplore&mdash\;a JAVA application for presentin
 g\, visualizing and interrogating large complex networks of interactions. I
  illustrate the use of HdBCS and GraphExplore to efficiently mine across mu
 ltiple microarray data sets. \n\n Next I develop a comprehensive framework 
 for combining ordered categorical and continuous covariates into parsimonio
 us predictive models for categorical variables with two or more levels. I e
 mphasize the importance of parallel computing in exploring huge spaces with
  thousands of covariates. The final example of the talk involves the identi
 fication of multivariate patterns of association among gene expression prof
 iles\, SNPs and clinical data that are predictive of atherosclerosis burden
  in human target tissues.
SUMMARY:High-dimensional Structural Learning and its Applications to Data F
 usion: Microarray Experiments\, Gene Expression Profiles\, SNPs and Clinica
 l Data
DTSTART:20060307T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:102
SEQUENCE:0
DTEND:20060221T160000
UID:2009-11-17T20:55:59-08:00_653425656@limen.stat.ucla.edu
DESCRIPTION:In this talk\, a general class of models referred to as perturb
 ation models are introduced. These models are described by an underlying "n
 ull" model that accounts for most of the structure in data while a perturba
 tion accounts for possible small localized departures. For instance\, in th
 e context of finite mixture models\, the null model represents a mixture wi
 th m components and the perturbation model represents additional components
 . In the spatial scan process context\, the null density accounts for the b
 ackground or noise whereas the perturbation searches for an unusual region 
 such as a tumorous tissue in mammography or a target in an image recognitio
 n problem. We derive a new test statistic for detecting the presence of per
 turbation and show that the asymptotic distribution of the test statistic i
 s equivalent to the supremum of a Gaussian process over a high-dimensional 
 manifold (e.g.\, curve\, surface etc.) with boundaries and singularities. A
  technique for approximating the quantiles of the  test statistic via the H
 otelling-Weyl volume-of-tube formula is presented. \n\n Fitting mixture mod
 els and performing statistical inference on the results is an important but
  a very challenging problem. A long-pending fundamental question is: how ma
 ny mixture components?  The asymptotic null distribution of the likelihood 
 ratio test statistic is highly complex and very difficult to simulate from 
 in practice. Building on the perturbation theory\, inferential methods are 
 developed to address the problem of testing for an arbitrary number of comp
 onents from smooth families of distributions\, including multivariate mixtu
 res. The resulting theory has broad applications including astronomy\, astr
 ophysics\, particle physics\, bioinformatics and genetics. We illustrate th
 e theory in the context of a model problem from high-energy particle physic
 s wherein the goal is to distinguish a signal from random fluctuation in da
 ta with a high probability. More information on the particle physics and ot
 her application problems is available at http://stat.case.edu/~pillar/PRL/P
 RL.htm
SUMMARY:Inference in Perturbation Models\, Mixtures and Spatial Scan Proces
 s
DTSTART:20060221T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:104
SEQUENCE:0
DTEND:20060207T160000
UID:2009-11-17T20:55:59-08:00_887020136@limen.stat.ucla.edu
DESCRIPTION:A central theme of research on human development and psychopath
 ology is whether a therapeutic intervention or a turning point event\, such
  as a family break-up\, alters the trajectory of the behavior under study. 
 This paper lays out and applies a method for using observational longitudin
 al data to make more confident causal inferences about the impact of such e
 vents on developmental trajectories. The method draws upon two distinct lin
 es of research: Work on the use of finite mixture modeling to analyze devel
 opmental trajectories and work on propensity scores. The essence of the met
 hod is to use the posterior probabilities of trajectory group membership fr
 om a finite mixture modeling framework to create balance on lagged outcomes
  and other covariates established prior to t for the purpose of inferring t
 he impact of first-time treatment at t on the outcome of interest. The appr
 oach is demonstrated with an analysis of the impact of gang membership on v
 iolent delinquency based on data from a large longitudinal study conducted 
 in Montreal.
SUMMARY:Causal Inference with Longitudinal Data: A Case Study of Gang Membe
 rship and Violent Delinquency
DTSTART:20060207T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:114
SEQUENCE:0
DTEND:20051101T160000
UID:2009-11-17T20:55:59-08:00_116678113@limen.stat.ucla.edu
DESCRIPTION:The goal of the seminar is to present a method for dimension re
 duction in kernel nonparametric discriminant analysis. The procedure is bas
 ed on a general concept of separation of populations that leads to a new ch
 aracterization of the central subspace in discriminant analysis. Examples o
 f application and comparisons with other methods are also studied. \n\n \n\
 n \n\n Bio. Sketch. Santiago Velilla studied at Universidad Complutense de 
 Madrid (UCM)\, Spain. He obtained a degree in Mathematics in 1982\, and a P
 h.D. in Statistics and Operations Research in 1987. In 1988\, he became Ass
 ociate Profesor of the Department of Statistics and Operations Research of 
 the School of Mathematics of UCM. In 1990\, he joined Universidad Carlos II
 I de Madrid where\, from 1998\, he is Professor at the Department of Statis
 tics. His current research interests are Regression\, Time Series and Multi
 variate Analysis.
SUMMARY:Dimension Reduction in Nonparametric Discriminant Analysis
DTSTART:20051101T150000
DTSTAMP:20091117T205559
LOCATION:6627 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:118
SEQUENCE:0
DTEND:20051010T170000
UID:2009-11-17T20:55:59-08:00_500473272@limen.stat.ucla.edu
DESCRIPTION:Image analysis has gained significantly in quality over the las
 t decade by complex statistical models. Grouping algorithms based on local 
 histograms to represent image patches have shown satisfactory performance i
 n image segmentation\, i.e.\, if they are combined with feature selection. 
 I will discuss a nonparametric Bayesian approach to smooth image segmentati
 on where the algorithm determines the properties and the number of segments
  using a mixture of Dirichlet processes while simultaneously enforcing a Ma
 rkov Random Field constraint. In the second part of the talk I will discuss
  a structured statistical model for object recognition which is designed in
  the spirit of Geman's compositionality architecture. Feature histograms of
  local image patches are extracted to form "parts" which are then linked to
  combinations. Bags of combinations are then used to categorize images. Cro
 ss-validated test errors on the Caltech 101 database yield a categorization
  rate of 52 percent. \n\n Bio. sketch Joachim M. Buhmann is full Professor 
 for Computer Science (Information Science and Engineering) at ETH Zurich si
 nce October 2003. He received a diploma in physics in 1985 and a PhD in 198
 8\, both from the Technical University of Munich. Afterwards\, he spent thr
 ee years as research associate and research assistant professor at the Univ
 ersity of Southern California\, Los Angeles. In 1991 he joined the Lawrence
  Livermore National Laboratory in Livermore\, California. From 1992 until 2
 003 he was a professor for applied computer science at the University of Bo
 nn.
SUMMARY:Complex Statistical Models for Image Segmentation and Object Recogn
 ition
DTSTART:20051010T160000
DTSTAMP:20091117T205559
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:120
SEQUENCE:0
DTEND:20050607T160000
UID:2009-11-17T20:55:59-08:00_540332205@limen.stat.ucla.edu
DESCRIPTION:It is common in applied research to have large numbers of varia
 bles measured on a modest number of cases. Even with low rates of missingne
 ss on individual variables\, such data sets can have a large number of inco
 mplete cases.  Here we present new methods for handling missing continuousl
 y scaled items in multivariate data\, based on extracting common factors to
  reduce the number of covariance parameters to be estimated in a multivaria
 te normal model.  One technique is developed to handle cross-sectional data
  sets with general covariance patterns\, while the other is specifically ta
 ilored to accommodate longitudinal data with many-dimensional outcomes.  Si
 mulation studies compare the statistical properties of the methods with pot
 ential alternative approaches.  The methods are also illustrated in applied
  settings with over 100 variables\, one being an investigation of psycholog
 ical outcomes in a study of an emergency room intervention for adolescents 
 who attempted suicide and the other being a study of quality of life integr
 ated into a clinical trial on oral-surgery patients. \n\n (joint work with 
 Juwon Song\, Korea University\, and Jianming Wang\, Medtronic Vascular\, In
 c.)
SUMMARY:Bayesian Methods to Handle Missing Data in High-Dimensional Data Se
 ts using Factor Analysis Strategies
DTSTART:20050607T150000
DTSTAMP:20091117T205559
LOCATION:6229 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:124
SEQUENCE:0
DTEND:20050510T160000
UID:2009-11-17T20:55:59-08:00_555788936@limen.stat.ucla.edu
DESCRIPTION:As the rate of enrollment in lower division classes continue to
  grow at UCLA\, it has become more challenging to maintain the quality of i
 nstruction\, student-teacher interaction\, and constructive methods of stud
 ent evaluation. The College of Letters and Science at UCLA is looking into 
 blended instruction\, combining technology and customary teaching methods\,
  as a solution to this dilemma. To that end\, in 2004 the College awarded t
 hree departments including Statistics grants to conduct case studies to exa
 mine the potential of blended instruction as a possible solution to the pro
 blem described above. \n\n In Winter 2005 blended instruction was implement
 ed in Statistics 10\, which has the highest enrollment rate (around 1700-18
 00 per year) in the department. The major objectives were to introduce stat
 istics as a science of data\, maximize the role of students as active learn
 ers\, help the instructors and the TAs develop a better sense of the studen
 ts progress through on-line quizzes\, establish closer TA-student\, instruc
 tor-student\, and student-student contact\, and use assessment to enhance u
 pper level thinking and statistical thinking. \n\n The instructor and the t
 wo teaching assistants who conducted the blended instruction case study wil
 l share their experiences in this seminar.
SUMMARY:An Instructor and Two Teaching Assistants Share Their Experiences W
 ith Blended Instruction
DTSTART:20050510T150000
DTSTAMP:20091117T205559
LOCATION:9413 Boelter Hall
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:128
SEQUENCE:0
DTEND:20050413T160000
UID:2009-11-17T20:55:59-08:00_874407144@limen.stat.ucla.edu
DESCRIPTION:We consider the problem of estimating a vector of location para
 meters restricted to a cone in the presence of an unknown scale parameter. 
 Examples include estimating the location vector when the parameter ordering
  is known (or partially known)\, or when it is known that all means are pos
 itive. A standard estimator when sampling from a multivariate normal distri
 bution is the MLE. In the unrestricted case\, this estimator is the vector 
 of sample means\, which is dominated (in three and higher dimensions) by th
 e James-Stein estimator among others. We study improvements of the James-St
 ein type for the restricted parameter case in the general setting of a sphe
 rical symmetry location family. The development is based on a general resul
 t that shows how improved estimators of location in the multivariate normal
  case with known scale can be extended to give improved estimators in the g
 eneral spherically symmetric case with unknown scale.
SUMMARY:Improved Estimation of Restricted Parameters
DTSTART:20050413T150000
DTSTAMP:20091117T205559
LOCATION:5200 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:130
SEQUENCE:0
DTEND:20050308T160000
UID:2009-11-17T20:55:59-08:00_208460019@limen.stat.ucla.edu
DESCRIPTION:Clustering is a popular tool for the exploratory analysis of ge
 ne expression data. However\, the specific structure and characteristic of 
 gene expression data presents challenges for most off-the-shelf clustering 
 routines. In this talk we discuss some inherent difficulties in clustering 
 microarray data. We also present two new statistical methodologies for clus
 tering. \n\n DDclust is a clustering method based on the concept of data de
 pth. It incoorporates clustering and cluster validation into one step\, and
  is geared toward discovering clusters of different scale - something that 
 many of the most commonly used algorithms cannot. \n\n PClust is a method w
 hich allows for the simultaneous selection and clustering of genes based on
  data from a replicate microarray experiment. Most clustering methods do no
 t take the experimental setup into account and relies on arbitrary measures
  of similarity. PClust is based on statistical hypothesis testing and gene-
 gene similarity is defined through the outcome of the test. We demonstrate 
 how PClust adapts to data availability\, and how the simultaneous selection
  and clustering approach can lead to significant increases in power of sele
 ction.
SUMMARY:Clustering Gene Expression Data - Some Statistical Approaches
DTSTART:20050308T150000
DTSTAMP:20091117T205559
LOCATION:5128 Math Sciences Bldg.
END:VEVENT
BEGIN:VEVENT
X-SEMINAR-ID:134
SEQUENCE:0
DTEND:20050215T160000
UID:2009-11-17T20:55:59-08:00_633218436@limen.stat.ucla.edu
DESCRIPTION:Industrialized fishing is dramatically reducing the stock of pr
 edatory fish throughout the oceans of the world. Large-scale commercial fis
 hing affects not just the target species but other species that become the 
 'bycatch'. The impact on dolphin populations of commercial fishing for tuna
  is perhaps the most visible illustration and has been the subject of a U.S
 . National Research Council committee report. Over the past decade\, intern
 ational cooperation to reduce dolphin mortality has led to efforts to monit
 or tuna fishing practices and penalize offenders. \n\n We have recently acq
 uired data from Inter-American Tropical Tuna Commission (ITTC)\, which over
 sees international purse-seine fishery for tuna in the eastern Pacific Ocea
 n. We have begun working with the ITCC to develop a methods to predict when
  dolphin are put at risk and to flag sets in which the onboard observer res
 ponsible for data collection may have systematically underreported the numb
 er of dolphin killed to prevent sanctions from being applied. The number of
  observations is very large and the number of underreports is presumably ve
 ry small. \n\n In this talk\, the nature of the problem will be considered 
 and some preliminary results reported. The factors that put dolphin at risk
  are often reasonably clear and some can b