Synopsis Data Structures for Approximate Query Answering

Phillip B. Gibbons

"Synopsis data structures" for massive data sets seek to summarize a
data set using substantively less space than needed to store the full
data.  Various statistical and algorithmic techniques are used to
compute and keep up-to-date the synopses, with the goal of providing
highly-accurate approximate answers to (typically aggregate) queries
on the full data set.  A synopsis data structure offers the following
advantages over traditional, linear space data structures: (a) it may
reside in main memory, enabling query responses and data structure
updates that avoid disk accesses altogether, (b) it can be transmitted
remotely at minimal cost, (c) it has minimal impact on the overall
storage costs of a system, (d) it leaves space in the memory for other
processing (available main memory is a precious resource for
applications on massive data sets), and (e) it can serve as a small
surrogate for data sets that are currently expensive or impossible to
access.

The design and analysis of effective synopsis data structures offer many algorithmic and statistical challenges. We discuss several concrete examples of synopsis data structures, and describe fast algorithms for keeping them up-to-date in the presence of online updates to the data sets. Finally, we describe the Approximate Query Answering (Aqua) prototype, the first system to employ synopsis data structures in order to provide very fast approximate answers to queries on large databases.

The talk will be held at 10:30am  on Thu, Jul 15 in Room MH 2D-103.

Sponsor: José Pinheiro, MH 2C-258, x2390

[ Return to seminar list ]

Statistics and the pricing of options

Per Mykland

There is a substantial body of theory on how to price and
hedge options when the probabilistic model for the underlying
security is known and continuous. There is also a big
literature on statistical inference in such models. It is
not clear, however, how statistical uncertainty can be incorporated
into prices. Indeed, many practitioners do not rely on statistics
in their trading.

In this talk, we discuss ways of bridging this gap. It will be shown that confidence and prediction intervals can typically be converted to worst case hedging strategies. Such strategies can then be used to set up reserve requirements, or directly for trading. The procedure in principle divorces the statistical and the trading problem, but we shall also see that the market structure has implications for how to carry out the statistical analysis.

In addition to providing a link between statistics and options trading, the device also reduces one's dependence on expert judgement in pricing this type of instrument, something which is greatly to be desired, both to improve management oversight and government regulation.

The talk will be held at 10:30am on Thu, Oct 7 in Room MH 2D-101 .

Sponsor: Bin Yu, MH 2C-260, x3257

[ Return to seminar list ]

NPMLE Based on Coarse Data in sigma-VC Helly Classes

Cun-Hui Zhang

We consider nonparametric maximum likelihood estimators of a probability
measure based on coarse data.  Coarsening is a general model of incomplete
data which includes grouping and censoring with measurable sets. Asymptotic
consistency of nonparametric maximum likelihood estimators is established
under nearly sharp conditions in several modes of convergence. The results
cover the Turnbull (1976) set-censoring model and certain multivariate
extensions of it. The proofs are based on a likelihood ratio inequality for
additive set functions in a field against measures, and on certain
characterizations of the nonparametric maximum likelihood estimators.

The talk will be held at 2:00pm  on Fri, Oct 08 in Room MH 2D-101.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

The Omegahat Statistical Software

John Chambers and Duncan Temple Lang

The Omega project for statistical computing is a joint effort to
develop the next generation of software for statistics.  Participants
include many of the leading researchers in statistical computing.  The
first release of the open-source Omegahat software was announced in
September.  The current release is not intended as a statistical
system, but rather as an environment for experimenting with some
important new directions in statistical computing.  Most of the
software is written in Java, including an interactive environment,
plus interfaces from existing systems to Java and CORBA.

This is the first of a proposed series of informal seminars describing the new software. The first talk will give an overview and some examples of the new approach, with the main goal of encouraging comments and participation from anyone interested in statistical software. Later talks will include discussion of the interactive environment, inter-system interfaces, and graphics.

The talk will be held at 10:30am on Tue, Oct 12 in Room MH 2D-101.

[ Return to seminar list ]

Stochastic Optimization Methods for Fitting Polychotomous Regression (POLYCLASS) and Feed-Forward Neural Network (NN) Models

Charles J. Stone

POLYCLASS uses adaptively selected linear splines and their tensor products
to fit the various canonical regression functions appearing in a model for
a polychotomous response variable and one or more predictors. Classical
numerical methods for obtaining the maximum likelihood estimates are feasible
when the numbers of classes, cases, and features are not too large, but
become impractical beyond a certain point.  We gain further insight into the
fitting of large POLYCLASS models by simultaneously considering the fitting
of large NN models. In this combined setting, the stochastic gradient method,
as used in the on-line version of the backpropagation method for fitting NN
models, and a stochastic version of the conjugate gradient method for fitting
such models emerge as being computationally attractive. In particular, these
stochastic methods are successfully applied to the fitting of POLYCLASS and
NN models in the context of a phoneme recognition problem involving 45
phonemes, 81 features, 150,000 cases in the training sample, up to 1000 basis
functions and 44,000 parameters for POLYCLASS, and up to 800 hidden nodes and
about 100,000 parameters for NN.

The talk will be held at 10:30am on Thu, Oct. 28 in Room MH 2D-101 .

Sponsor: Mark Hansen, MH 2C-283, 3869

[ Return to seminar list ]

The Omegahat Environment

John Chambers and Duncan Temple Lang

This talk introduces the Omegahat interactive environment, a
Java-based computing environment.  The environment provides
interactive evaluation of a language that extends Java, but also has a
functional form similar to S, with user-defined functions, classes,
and methods.  Unlike S and similar current systems, however, Omegahat
is component-based: All the components, such as the parser and
evaluator, are fully extensible within the environment itself.
Java-based features, such as threading and event management, also
provide new approaches to statistical software design.

The talk will describe the first version of the environment, suitable for experimental use as part of research in statistical software. This is one of a series of talks introducing the Omega project for statistical computing.

The talk will be held at 10:30am on Wed, Nov 03 in Room MH 2D-101.

[ Return to seminar list ]

Theory and Practice of Boosting Robert Schapire

Boosting is a general method for producing a very accurate
classification rule by combining rough and moderately inaccurate
"rules of thumb."  While rooted in a theoretical framework of machine
learning, boosting has been found to perform quite well empirically.
In this talk, I will introduce the boosting algorithm AdaBoost, and
explain the underlying theory of boosting, including an explanation of
why boosting often does not suffer from overfitting.  I also will
describe some recent applications of boosting.

The talk will be held at 10:30am on Tue, Nov 09 in Room MH 2D-101.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Inter-System Interfaces in Omegahat

John Chambers and Duncan Temple Lang

The Omegahat software includes a set of inter-system interfaces; that
is, facilities to allow users of one system to access functionality in
other languages and applications, and also provide services from their
environment to others.

The first uses dynamic CORBA facilities to provide access to and from R, S and Omegahat and other applications, and to support high performance distributed computing. The second provides within-process interfaces from S and R to Java, using tools from the Omegahat environment.

Both interfaces use a notion of object reference and callback to allow arbitrary data structures to be made available within their native language. The approach is motivated by the requirements of CORBA, but is helpful for the other interfaces as well.

This is the third in a series of talks describing current research in the Omega project, a joint project aimed at designing the next generation of software for statistical computing.

The talk will be held at 10:30am on Wed, Nov 10 in Room MH 2D-101.

[ Return to seminar list ]

The Omegahat Graphics Engine

Paul Murrell

The Omegahat software includes a Graphics module that provides a
powerful idiom for statistical graphics, made possible in part by the
power of the underlying Java graphics model and in part by a
``graphical engine'' with an approach to the elements of the graphics
objects that makes for flexible, hierarchical design.
In particular, the graphics engine makes use of general concepts of
rectangular regions and graphical objects that, within the overall
approach of Java and Omegahat, should be useful in the design of new
graphical software for statistics.

This talk describes the initial design and current contents of the graphical engine.

This is the fourth in a series of talks describing current research in the Omega project, a joint project aimed at designing the next generation of software for statistical computing.

The talk will be held at 10:30am on Wed, Nov 17 in Room MH 2D-101.

Sponsor: Duncan Temple Lang, MH 2C-259, x3217

[ Return to seminar list ]

Variable Length Markov Chains and Nonlinear Time Series

Peter Buhlmann

We present some modeling techniques for categorical and nonlinear
(real- or vector-valued) time series.

Variable length Markov chains (VLMC) are attractive tree structured
Markov models for finite alphabet categorical time series. They can be
efficiently estimated with a version of Rissanen's context
algorithm. Advantages of VLMC's are their flexibility, interpretation
(which we demonstrate on a DNA data set) and performance.

Carrying over the VLMC idea to real- or even vector-valued time series
is a much more difficult task due to much bigger state spaces in the
Markovian world. We propose: (i) quantization and VLMC modeling and;
(ii) dynamic combination of models with VLMC mixture weights. The
latter proposal (ii) allows for non-Markovian and likelihood based

The talk will be held at 10:30am on Thu, Nov 18 in Room MH 2D-101.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Statistical Representations of Music and their Applications

Rafael Irizarry

Statistics has been applied in various ways to music. For example,
various stochastic techniques have been applied in
composition. Stochastic techniques have also been used in forecasting
unfinished works.  The spectral properties of different musical
signals have been studied for the possibility of it being so called
1/f noise.  For this talk the particular application that will be
examined in detail is the analysis of sound signals produced by
musical instruments. In this field, statistical techniques have been
used, for example, to separate the signals into deterministic and
stochastic parts and to deconstruct the deterministic part into
harmonic components. Residual analysis by ear is used to determine
goodness of fit.

The talk will be held at 10:30am on Thu, Dec 09 in Room MH 2D-101.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Statistical Prediction for Low-Bit Rate/High Quality Sound Coding

Dawei Huang and Bin Yu

The demand is increasing rapidly for broadcasting via Internet or
wireless.  Mixed (music and speech) digital signals need to be
transmitted at low bit rates but with high quality. In this talk, we
investigate a statistical prediction approach in a new coding paradigm
that separates irrelevancy and redundancy reductions to allow
individual fine tunings of each part.  The irrelevancy reduction is
carried out via an adaptive prefilter based on a psychoacoustic
model. The quantized prefiltered signal is then subject to our
statistical (integer) prediction whose (integer) residuals are
entropy-coded to achieve redundancy reduction. The decoded signal is
passed through a post-filter to obtain the reconstructed signal.

Our first phase prediction is the LMS (Least Mean Square) algorithm,
the second applies LMS again to the residuals of the first phase, and
the third does the same to the residuals of the second phase. Or we
cascade LMS three times to obtain three predictors. Our final
prediction scheme is either the median or the weighted average of the
three predictors. The weights are chosen according to the Predictive
MDL Principle. Both final predictors achieve similar performance on
test samples, but the latter is easier to analyze analytically.

The combined coder achieves a low bite rate of 1.5-2 bits per sample
(32kHZ sampling rate) with an FM quality reconstructed signal.  In
progress is the statistical prediction between channels in stereo
signals.

(The talk material is part of a project jointly with G. Schuller and
B. Edler on coding mixed signals.)

The talk will be held at 10:30am on Tue, Feb 08 in Room MH 2D-101.

[ Return to seminar list ]

Run Related Probability Functions Applied to Quality Control

Galit Shmueli

Various statistical procedures that are used to control quality
include switching or stopping rules. These rules are usually based on
a previous results, and involve the concept of runs. A run is a
sequence of identical events, such as a sequence of successes in a
slot machine. However, waiting for a run to occur in not merely a
superstitious act. In quality control, as in many other fields
(e.g. reliability of engineering systems, DNA sequencing,etc.), the
concept of runs is widely applied as the underlying basis for many
rules.

Run rules are usually very intuitive and simple to apply (e.g. "use
reduced inspection following a run of 5 acceptable batches"). In fact,
they were designed according to empirical rather than probabilistic
considerations.  Therefore, there is a need to investigate their
theoretical properties and to assess their performance in light of
practical requirements. In order to investigate the properties of such
systems their complete probabilistic structure should be revealed.

Various authors addressed the occurrence of runs from a theoretical
point of view, with no regard to the field of quality control. The
main problem has been to specify the exact probability functions of
variables which are related to runs. This problem was approached from
different methods, some of them leading to expressions for the pdf.

We present an alternative method for obtaining the pdf of run-related
variables (including order k variables). Our method is based on
recurrence relations, which lead to an expression for the probability
generating function of many run-related variables. We then use
the special form of the generating function to derive the pdf by
partial fraction expansion (Feller, 1968), with the aid of efficient
numerical methods.

Finally, we apply our results to study systems which are based on
runs, and in particular to the switching rules between levels of
inspection in acceptance sampling and to the Western Electric
sensitizing rules for Shewhart control charts. For the latter system
we present an Internet application, which gives the exact run-length
distribution, when various run-rules are applied to a Shewhart
control chart. Other applications (to continuous sampling and system
reliability) will be mentioned as well.

The talk will be held at 10:30am on Thu, Feb 17 in Room MH 2D-101.

Sponsor: José Pinheiro, MH 2C-258, x2390

[ Return to seminar list ]

Some Statistical Issues in Molecular Biology and Genetics

Hongyu Zhao

With the rapid progress in the Human Genome Project, it is becoming
a reality to study all genetic variations in humans simultaneously.
These technological advances have created both an exciting and
challenging opportunity for statisticians to develop novel
statistical tools to take advantage of the large amount of
biological information generated from this current biological
evolution. In this talk, I will discuss two topics that have
recently drawn a lot of attention. One is the use of large numbers
of genetic markers to locate disease genes, and the other is
the use of microarray technologies to study gene expression patterns
and genetic networks. For each topic, I will describe the
biological background, types of genetic data that are
utilized to address the biological questions, limitations of the
current statistical methods, and the novel statistical
methods under development.

The talk will be held at 1:30pm on Fri, Mar 10 in Room MH 2D-101.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Bayesian Partition Models

David Denison

In this talk we propose a new Bayesian approach to data modelling
motivated by the difficulties encountered with some tree-based
methods.  The Bayesian partition model constructs arbitrarily complex
response surfaces over the design space by splitting it into an
unknown number of disjoint regions. Within each region the data is
assumed to be exchangeable and to come from some simple
distribution. Using conjugate priors the marginal likelihoods of the
models can be obtained analytically for any proposed partitioning of
the space, hugely simplifying the sampling algorithm required to
simulate from the posterior of interest.  By example we shall show how
the partition model can be used for a wide variety of problems
including regression, classification and disease mapping.

The talk will be held at 10:30am on Tue, Mar 28 in Room MH 2D-101.

Sponsor: Mark Hansen, MH 2C-283, x3869

[ Return to seminar list ]

Some topics in the theory of statistical practice

Andrew Gelman

We report on some recent small research projects in which we have
tried to formalize some statistical procedures as they are actually
used.  As time permits, we shall present some of the following topics:

- Type S errors and multiple comparisons
- Choosing the treatment levels in an experimental design
- Survey weighting and poststratification
- Statistical interpretations of deterministic models
- Systematic artifacts in map displays

We then plan to discuss the challenges that arise in the formal study
of statistical methods.

The talk will be held at 2pm on Fri, Apr. 14 in Room MH 2D-101 .

Sponsor: Jin Cao, MH 2C-260, x2484

[ Return to seminar list ]

A Variational Approach to Estimation with Applications to Nonlinear Filtering

Sanjoy Mitter

In this talk I discuss the role of the Donsker - Varadhan Variational
Principle in Nonlinear Estimation.I show how this leads to a Stochastic
Control formulation for optimal filtering of diffusion processes.

The talk will be held at 10:30-11:30 on Wed, May 03 in Room MH 2C-528 .

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Networks for Networks: Mapping the Brain and the Internet using Graphical Statistical Methods

Robert Nowak

This talk describes a statistical analysis framework for inverse
problems involving Poisson data based on Bayesian tree-structured
models.  Such problems arise in medical image tomography and computer
network tomography applications.

For medical imaging, the tree-structured framework is based on a
multiscale analysis associated with recursive partitioning of the
underlying image intensity, a corresponding multiscale factorization
of the likelihood (induced by this analysis), and a choice of prior
probability distribution made to match this factorization by modeling
the ``splits'' in the underlying partition and reflect prior belief as
to the smoothness of the unknown intensity.  Adopting the
expectation-maximization (EM) algorithm for use in computing the
maximum a posteriori estimate corresponding to the model, we find that
the model permits remarkably simple, closed-form expressions for the
EM update equations and leads to high quality image reconstructions.

Network tomography is a promising new technique for studying the
(internal) behaviour of large-scale computer networks based solely on
end-to-end measurements. Tree-structured models arise naturally here
due to the physical topology of computer networks.  Based on these
models, we demonstrate that it is only possible to resolve internal
losses if reasonable prior information and constraints are identified
and incorporated in an modeling/inference framework. Inference
calculation is carried out using a novel factor graph representation.
Simulation experiments demonstrate the potential of our new framework.

The talk will be held at 10:30-11:30 on Thu, May 11 in Room MH 2D-101 .

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Partial Regression and Correlation Curves Kjell Doksum

The conditional expected value of a response given a vector of
covariate values contain much of the information about the
relationship between the response and the covariates. However, this
regression surface cannot be estimated efficiently nor displayed when
the number of covariates is moderate or large. In this case, it is
useful to extract essential features of the regression surface that
can be estimated efficiently and that can be displayed and interpreted
easily. For this purpose, we explore local projections of the regression
surface onto the sample space of each covariate.  We estimate the
projections using locally linear methods and give significance plots
to display the regions where the response is significantly correlated
with each covariate. We propose to select the bandwidth by
maximizing the power of the procedure.  This is joint work with Alex
Samarov and Chad Schafer

The talk will be held at 10:30am on Thu, May 18 in Room MH 2C-528.

Sponsor: José Pinheiro, MH 2C-258, x2390

[ Return to seminar list ]

Wireless Communications as a Statistical Problem

Vincent Poor

Wireless communications is one of the most active areas of technology
development today. Driven largely by the convergence of wireless and
the Internet, the exponentiating demand for wireless services is
presenting enormous challenges to researchers and system designers in
this area.  The drive to push system capacity, quality of service, and
mobility, well beyond their preconceived limits, is being enabled by
striking innovations in signal processing methods and
technology. These innovations include direct developments, such as
multimedia compression, multiuser detection, space-time processing and
coding, and turbo decoding, as well as indirect developments, such as
the use of signal processing techniques to push practical
micro-lithography beyond its current limits.

Many of these key signal processing problems arising in wireless
technology are essentially statistical problems, whose solutions lie
in the the methods of statistical inference.  This talk will discuss
this aspect of signal processing for wireless communications.  After
an overview of the wireless area and the role of signal processing in
its advancement, the talk will focus on several research problems
arising in the field for which statistical methods hold the key to
solution.

The talk will be held at 10:30am on Thu, May 25 in Room MH 2D-103.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

Some Statistical Methods for Blind Deconvolution in Data Communications

Ta-Hsin Li

With the ever-increasing demand for high speed in single/multiple user
wired/wireless data communications, the capacity of transmission
channels are being pushed to their physical limits. Under such
environments, the transmitted signals suffer from severe distortions
due to inter-symbol interference (ISI) and/or inter-channel
interference (ICI).  To enhance the transmission performance,
engineers have been interested in designing filters to remove the
interferences from the received data.  This requires sophisticated
statistical methods, especially when the training data are not
available (the 'blind' deconvolution problem). This talk will discuss
several statistical methods of blind deconvolution that exploit the
discretness of the input signal in data communications. In particular,
it will discuss the possible utility of the MCMC (Monte Carlo Markov
Chain) technique to help compute a Bayesian estimator for joint
channel and signal estimation.

The talk will be held at 1:30pm on Thu, Jun 01 in Room MH 2A-435.

Sponsor: Bin Yu, MH 2C-261, x3257

[ Return to seminar list ]

The multivariate L_1 median: a new algorithm and a new concept of data-depth in mutivariate data-analysis

Yehuda Vardi

The L_1 median is a useful multivariate statistical tool. To
mathematicians it's known as the The Fermat-Weber location problem
(AKA
the  Euclidean Steiner problem). It has been around for nearly
three centuries, and the most widely used algorithm for it is the well
known Weiszfeld (1938) algoritm, which has certain convergence
problems.
We propose a new algorithm (equally simple and elegant) which is
guarantied to converge monotonically to the desired median.
We further define a new concept for 'data-depth' of multivariate data
(or
distribution) and show that for the L_1 and several other  medians,
this
new definition results in   simple closed-form formulae which are easy
to compute in any dimension, making the new concept a practical
multivariate data-analytic tool.

The talk will be held at 10:30am on Thu, July 27 in Room MH 2D-103 .

Sponsor: Jin Cao, MH 2C-260, x2484

[ Return to seminar list ]

Statistical Modeling of UNIX Users and Processes with Application to Computer Intrusion Detection

Wen-Hua Ju

 High-order Markov chain and rarity models are proposed for
  profiling UNIX users and processes to identify their
  "signature behaviors". Sequences of shell commands and system
  calls are employed as the audit data stream for users and
  processes, respectively. The goal of monitoring is to detect
  potential intrusions by discovering irregularities in their
  behaviors.

  Starting from a high-order Markov chain model, the first method
  evolves to a hybrid model which is based mostly on the Markov
  chain element and occasionally on a statistical-independence
  model. A procedure driven by Maximum Likelihood (ML)
  considerations is devised to estimate the model parameters as the
  profile of normal behavior. The formal ML estimates are
  numerically intractable, but the ML-optimization problem can be
  substituted by a linear inverse problem with positivity constraint
  (LININPOS), for which the EM algorithm can be used as an equation
  solver to produce an approximate ML-estimate. The second method
  focuses on the ``rarity'' of short sequences of audit data which
  is measured by their relative usage frequencies. Individual
  commands (or system calls) as well as the transitions between them
  are taken into account to formulate scores. It requires little
  computing resource and can be easily implemented.

  The intrusion detection system works by comparing a sequence of
  audit data to the corresponding estimated signature behaviors in
  real time through statistical hypothesis testing. A form of
  likelihood-ratio test is devised. In the case of user profiling,
  it is used to detect if a given sequence of commands is from the
  proclaimed user, with the alternative hypothesis being a
  masquerader. In the case of processes modeling we test if a target
  system-call trace is part of the execution of an intrusion
  program. Real-life data collected from AT&T Labs - Research and
  publicly available data from the university of New Mexico are
  utilized to assess the proposed approaches. The experiment results
  indicate that our models hold promise and opens the door for many
  exciting modifications and extensions for computer intrusion
  detection.

The talk will be held at 10:30am on Thu, Aug 3 in Room MH 2D-103 .

Sponsor: Jose Pinheiro, x2390, room 2C-258

[ Return to seminar list ]

mmvz: An Axis-based Library for Interactive Data Visualizatio

Graham Wills

Novel methods for interactive graphic representations of information
have been described in numerous fields, and implementations of them
are available for different platforms as monolithic applications.
One of the problems facing researchers into fundamental techniques
for interactive visual display is the lack of a core library
designed to help build interactive visualizations. The previous
work of our Data Visualization group at Indian Hill allowed
us to build useful visualization tools in C++, and this has been
commercialized as the Visual Insights venture. Our latest effort
has been to build a multi-platform library that others can use
to build general interactive view. We call this library mmvz; it is
a Java library which was designed using a new Software Engineering
technique called "Commonality Analysis" and is being used to
implement new converged services for voice and data that will become
a Lucent product, as well as being trialed with other BUs. In this
talk I'll explain the Commonality Analysis we did and the
fundamental results we obtained (including the axis first principle),
describe the current status of the library and demo a few
applications.

The talk will be held at 10:30am on Thu, Oct 12 in Room MH 2D-101 .

Sponsor: David James, MH 2C-253, x3082

[ Return to seminar list ]

Sonification for Exploratory Data Analysis

Tomas Hermann

Sonification is the presentation of data by sound. This can be used as an
alternative or an extension to data visualization, to  discern structures and to
find  regularities in large
high-dimensional data sets.

This talk focusses on Model-Based Sonification, which is a concept to render
sonifications for arbitrary high-dimensional data sets. Sonification Models
facilitate the interpretation of the sound, provide new mechanisms to interact
with data and can be designed so that the sound reveals certain structural
properties in the data, like clustering, class borders or intrinsic
dimensionality. Some Sonification Models, including "Principal Curve
Sonification', 'Particle Trajectories in a Data Potential' and 'Data Sonograms'
are presented and sound examples are given.

Besides that, NEO/Nst, a graphical simulation and programming environment
developed in Bielefeld, is presented. NEO/Nst is especially well suited for
applications in the domain of artificial neural networks and datamining and is
also used to render sonifications.

The talk will be held at 10:30am on Thu, Oct 26 in Room MH 2D-101 .

Sponsor: Mark Hansen, MH 2C-283, x3869

[ Return to seminar list ]

Thinking Statistically about Optimization: A Bayesian Quasi-Newton Method

Chuanhai Liu

Quasi-Newton methods for unconstrained optimization have been developed
and scrutinized for half a century. The Broyden-Fletcher-Goldfarb-Shanno
(BFGS) method is commonly recognized as being stable and efficient.  In
the last decade, Broyden updates with negative parameters have sometimes
been found to result in more efficient algorithms. Using a canonical
representation of the current Hessian approximation, we first introduce
an intuitive method for updating the Hessian approximation. The new
method is in the Broyden  family and has a negative Broyden parameter.
The canonical representation also provides a  way to understand Hessian
update formulas for methods in the Broyden family, such as BFGS and
Davidon-Fletcher-Powell (DFP). In order to improve efficiency of the
intuitive method, we  model the uncertainty of the Hessian in directions
orthogonal to the current search direction. This results in a Bayesian
algorithm that automatically satisfies the quasi-Newton condition and
typically converges faster than both the BFGS and the intuitive method.
Our Bayesian formulation also explains why  BFGS is optimal near
convergence and how negative Broyden parameters can reduce the number of
iterations until convergence.

The talk will be held at 10:30am on Thu, Nov 30 in Room MH 2D-101 .

Sponsor: Jin Cao, MH 2C-260, x2484

[ Return to seminar list ]