NIPS*92 WORKSHOP PROGRAM

[Steve Hanson]

			NIPS*92 WORKSHOP PROGRAM


For Further information and queries on workshop please
respond to WORKSHOP CHAIRPERSONS listed below

=========================================================================
Character Recognition Workshop

Organizers: C. L. Wilson and M. D. Garris, NIST

Abstract:
In order to discuss recent developments and research in OCR technology,
six speakers have been invited to share from their organization's own
perspective on the subject. Those invited, represent a diversified
group of organizations actively developing OCR systems. Each speaker
participated in the first OCR Systems Conference sponsored by the Bureau
of the Census and hosted by NIST. Therefore, the impressions and results
gained from the conference should provide significant context for
discussions.

Invited presentations:
C. L. Wilson, NIST, "Census OCR Results - Are Neural Networks Better?"
T. P. Vogl, ERIM, "Effect of Training Set Size on OCR Accuracy"
C. L. Scofield, Nestor, "Multiple Network Architectures for Handprint
                         and Cursive Recognition"
A. Rao, Kodak, "Directions in OCR Research and Document Understanding
                at Eastman Kodak Company"
C. J. C. Burges, ATT, "Overview of ATT OCR Technology"
K. M. Mohiuddin, IBM, "Handwriting OCR Work at IBM Almaden Research Center"
=========================================================================
Neural Chips: State of the Art and Perspectives.

Organizer: Eros Pasero   pasero at polito.it

Abstract:
We will encourage lively audience discussion of important issues
in neural net hardware, such as:
- Taxonomy: neural computer, neural processor, neural coprocessor
- Digital vs. Analog: limits and benefits of the two approaches.
- Algorithms or neural constraints?
- Neural chips implemented in universities
- Industrial chips (e.g. Intel, AT&T, Synaptics)
- Future perspectives

Invited presentations: TBA
=========================================================================
Reading the Entrails: Understanding What's Going On Inside a Neural Net

Organizer: Scott E. Fahlman, Carnegie Mellon University
	   fahlman at cs.cmu.edu

Abstract:
Neural networks can be viewed as "black boxes" that learn from examples,
but often it is useful to figure out what sort of internal knowledge
representation (or set of "features") is being employed, or how the inputs
are combined to produce particular outputs.  There are many reasons why we
might seek such understanding: It can tell us which inputs really are
needed and which are the most critical in producing a given output.  It can
produce explanations that give us more confidence in the network's
decisions.  It can help us to understand how the network would react to new
situations.  It can give us insight into problems with the network's
performance, stability, or learning behavior.  Sometimes, it's just a
matter of scientific curiosity: if a network does something impressive, we
want to know how it works.

In this workshop we will survey the available techniques for understanding
what is happening inside a neural network, both during and after training.
We plan to have a number of presenters who can describe or demonstrate
various network-understanding techniques, and who can tell us what useful
insights were gained using these techniques.  Where appropriate, presenters
will be encouraged to use slides or videotape to illustrate their favorite
methods.

Among the techniques we will explore are the following: Diagrams of
weights, unit states, and their trajectories over time.  Diagrams of the
receptive fields of hidden units.  How to create meaningful diagrams in
high-dimensional spaces.  Techniques for extracting boolean or fuzzy
rule-sets from a trained network.  Techniques for extracting explanations
of individual network outputs or decisions.  Techniques for describing the
dynamic behavior of recurrent or time-domain networks.  Learning
pathologies and what they look like.

Invited presentations:
Still to be determined.  The workshop organizer would like to hear from
potential speakers who would like to give a short presentation of the kind
described above.  Techniques that have proven useful in real-world problems
are especially sought, as are short videotape segments showing network
=========================================================================
COMPUTATIONAL APPROACHES TO BIOLOGICAL SEQUENCE ANALYSIS--
   NEURAL NET VERSUS TRADITIONAL PERPECTIVES

Organizers: Paul Stolorz, Santa Fe Institute and Los Alamos National Lab
	    Jude Shavlik, University of Wisconsin.

Abstract:
There has been a good deal of recent interest in the use of neural
networks to tackle several important biological sequence analysis
problems. These problems range from the prediction of protein secondary
and tertiary structure, to the prediction of DNA protein coding regions
and regulatory sites, and the identification of homologies. Several
promising developments have been presented at NIPS meetings in the past
few years by researchers in the connectionist field.
Furthermore, a number of structural biologists and chemists have been
successfully using neural network methods.

The sequence analysis applications encompass a rather large amount of
neural network territory, ranging from feed forward architectures
to recurrent nets, Hidden Markov Models and related approaches.
The aim of this workshop is to review the progress made by these disparate
strands of endeavor, and to analyze their respective strengths and weaknesses.
In addition, the intention is to compare the class of neural network methods
with alternative approaches, both new and traditional. These alternatives
include knowledge based reasoning, standard non-parametric statistical
analysis,
Hidden Markov models and statistical physics methods.
We hope that by careful consideration and comparison of
neural nets with several of the alternatives mentioned above, methods can be
found which are superior to any of the individual techniques developed to date.
This discussion will be a major focus of the workshop, and we both anticipate
and encourage vigorous debate.

Invited presentations:
Jude Shavlik, U. Wisconsin: Learning Important Relations in Protein Structures
Gary Stormo, U. Colorado: TBA
Larry Hunter, National Library of Medicine:
     Bayesian Clustering of Protein Structures
Soren Brunak, DTH: Network analysis of protein structure and the genetic code
David Haussler, U.C. Santa Cruz: Modeling Protein Families with Hidden
     Markov Models
Paul Stolorz and Joe Bryngelson, Santa Fe Institute and Los Alamos:
     Information Theory and Statistical Physics in Protein Structures
=========================================================================
Statistical Regression Methods and Feedforward Nets

Organizers: Lei Xu, Harvard Univ. and Adam Krzyzak, Concordia Univ.

Abstract:
Feedforward neural networks are often used  for function
approximation, density estimation and pattern classification.
These tasks are also the purposes of statistical regression
methods. Some methods used in the literature of neural networks
and the literature of statistical regression are same, some are
different, and some have close relations. Recently, the
connections between the methods in the two literatures have been
explored from a number of aspects. E.g.,  (1) connecting feedforward
nets to parametric statistical regression for  theoretical studies
about multilayer feedforward nets;  (2) relating the
performances  of feedforward nets to the trade-off of bias and
variances in nonparameter statistics. (3) connecting Radial Basis
function nets to Nonparameter Kernal Regression to get  several
new theoretical results on approximation  ability, convergence
rate and receptive field size of  Radial Basis Function networks;
(4)  using VC dimension to study the generalization ability of
multilayer feedforward nets; (5)  using other statistical methods
such as projection pursuit,  cross-validation, EM algorithm, CART,
MARS for training feedforward nets. Not only in these mentioned
aspects there are still many interesting and open issues to be
further explored. But also,  in the literature of statistical
regression there are many other methods and theoretical results
on both nonparametric regression  and parameteric regression (e.g.,
L1 kernal estimation,  ..., etc).

Invited presentations:
Presentations will include arranged talks and submissions.  Submis-
sions can be sent to either of the two organizers by Email before
Nov.15, 1992. Each submission can be an abstract of 200--400 words.
=========================================================================
Computational Models of Visual Attention

Organizer: Pete Sandon, Dartmouth College

Abstract:
Visual attention refers to the process by which some part of the
visual field is selected over other parts for preferential processing.
The details of the attentional mechanism in humans has been the subject
of much recent psychophysical experimentation.
Along with the abundance of new data, a number of theories of attention
have been proposed, some in the form of computational models
simulated on computers.
The goal of this workshop is to bring together computational modelers
and experimentalists to evaluate the status of current theories
and to identify the most promising avenues for improving
understanding of the mechanisms and behavioral roles of visual
attention.

Invited presentations:
 Pete Sandon "The time course of selection"
 John Tsotsos "Inhibitory beam model of visual attention"
 Kyle Cave "Mapping the Allocation of Spatial Attention:
	    Knowing Where Not to Look"
 Mike Mozer "A principle for unsupervised decomposition and hierarchical
	     structuring of visual objects"
 Eric Lumer "On the interaction between perceptual grouping, object
   	     selection, and spatial orientation of attention"
 Steve Yantis "Mechanisms of human visual attention:
	       Bottom-up and top-down influences"
=========================================================================
Comparison and Unification of Algorithms, Loss Functions
and Complexity Measures for Learning

Organizers: Isabelle Guyon, Michael Kearns and Esther Levin, AT&T Bell Labs

Abstract:
The purpose of the workshop is an attempt to clarify and unify the
relationships
between many well-studied learning algorithms, loss functions, and
combinatorial
and statistical measures of learning problem complexity.

Many results investigating the principles underlying supervised learning from
empirical observations have the following general flavor: first, a "general
purpose" learning algorithm is chosen for study (for example, gradient descent
or maximum a posteriori). Next, an appropriate loss function is selected, and
the details of the learning model are specified (such as the mechanism
generating
the observations). The analysis results in a bound on the loss of the algorithm
in terms of a "complexity measure" such as the Vapnik-Chervonenkis dimension
or the statistical capacity.

We hope that reviewing the literature with an explicit emphasis on comparisons
between algorithms, loss functions and complexity measures will result in a
deeper understanding of the similarities and differences of the many possible
approaches to and analyses of supervised learning, and aid in extracting the
common general principles underlying all of them. Significant gaps in our
knowledge concerning these relationships will suggest new directions in
research.

Half of the available time has been reserved for discussion and informal
presentations.  We anticipate and encourage active audience participation.
Each discussion period will begin by soliciting topics of interest from the
participants for investigation. Thus, participants are strongly encouraged
to think about issues they would like to see discussed and clarified prior
to the workshop.  All talks will be tutorial in nature.

Invited presentations:
  Michael Kearns, Isabelle Guyon and Esther Levin:
	-Overview on loss functions
	-Overview on general purpose learning algorithms
	-Overview on complexity measures
  David Haussler: Overview on "Chinese menu" results
=========================================================================
Activity-Dependent Processes in Neural Development

Organizer: Adina Roskies, Salk Institute

Abstract: This workshop will focus on the role of activity in setting
up neural architectures. Biological systems rely upon a variety of
cues, both activity-dependent and independent, in establishing their
architectures. Network architectures have traditionally been
pre-specified, but it is ongoing construction of architectures may
endow networks with more computational power than do static
architectures.  Biological issues such as the role of activity in
development, the mechanisms by which it operates, and the type of
activity necessary will be explored, as well as computational issues
such as the computational value of such processes, the relation to
hebbian learning, and constructivist algorithms.

Invited presentations:
	General Overview (Adina Roskies)
	The role of NMDA in cortical development (Tony Bell)
	Optimality, local learning rules, and the emergence of function in a
		sensory processing network (Ralph Linsker)
	Mechanisms and models of neural development through rapid
		volume signals (Read Montague)
	The role of activity in cortical development and plasticity
		(Brad Schlaggar)
	Computational advantages of constructivist algorithms (Steve Quartz)
	Learning, development, and evolution (Rik Belew)
=========================================================================
DETERMINSTIC ANNEALING AND COMBINATORIAL OPTIMIZATION

Organizer: Anand Rangarajan, Yale Univ.

Abstract: Optimization problems defined on ``mixed variables'' (analog
and digital) occur in a wide variety of connectionist applications.
Recently, several advances have been made in deterministic annealing
techniques for optimization. Deterministic annealing is a faster and
more efficient alternative to simulated annealing. This workshop
will focus on several of these new techniques (emerging in the last
two years). Topics include improved elastic nets for the traveling salesman
problem, new algorithms for graph matching, relationship between
deterministic annealing algorithms and older, more conventional techniques,
applications in early vision problems like surface reconstruction, internal
generation of annealing schedules, etc.

Invited presentations:
	Alan Yuille, Statistical Physics algorithms that converge
	Chien-Ping Lu, Competitive elastic nets for TSP
	Paul Stolorz, Recasting deterministic annealing as constrained
	optimization
	Davi Geiger, Surface reconstruction from uncertain data
	on images and stereo images.
	Anand Rangarajan, A new deterministic annealing algorithm for
	graph matching
=========================================================================
The Computational Neuron

Organizer: Terry Sejnowski, Salk Institute (tsejnowski at ucsd.edu)

Abstract:
Neurons are complex dynamical systems.  Nonlinear properties arise
from voltage-sensitive ionic currents and synaptic conductances; branched
dendrites provide a geometric substrata for synaptic integration and learning
mechanisms.  What can subthreshold nonlinearities in dendrites be used to
compute?  How do the time courses of ionic currents affect synaptic
integration and Hebbian learning mechanisms?  How are ionic channels in
dendrites regulated?  Why are there so many different types of neurons?
These are a few of the issues that will we will be discussing.  In addition to
short scheduled presentations designed to stimulate discussion, we invite
members of the audience to present  one-viewgraph talks to introduce
additional topics.

Invited presentations:
	Larry Abbott - Neurons as dynamical systems.
	Tony Bell - Self-organization of ionic channels in neurons.
	Tom McKenna - Single neuron computation.
	Bart Mel - Computing capacity of dendrites.
=========================================================================
ROBOT LEARNING

Organizers: Sebastian Thrun (CMU), Tom Mitchell (CMU), David Cohn (MIT)

Abstract:
Robot learning has grasped the attention of many researchers over the
past few years. Previous robotics research has demonstrated the
difficulty of manually encoding sufficiently accurate models of the
robot and its environment to succeed at complex tasks. Recently a wide
variety of learning techniques ranging from statistical calibration
techniques to neural networks and reinforcement learning have been
applied to problems of perception, modeling and control.  Robot
learning is characterized by sensor noise, control error, dynamically
changing environments and the opportunity for learning by
experimentation.

This workshop will provide a forum for researchers active in the area
of robot learning and related fields.  It will include informal
tutorials and presentations of recent results, given by experts in
this field, as well as significant time for open discussion.  Problems
to be considered include: How can current learning robot techniques
scale to more complex domains, characterized by massive sensor input,
complex causal interactions, and long time scales?  How can previously
acquired knowledge accelerate subsequent learning? What
representations are appropriate and how can they be learned?

Invited speakers:
	Chris Atkeson
	Steve Hanson
	Satinder Singh
	Andrew W. Moore
	Richard Yee
	Andy Barto
	Tom Mitchell
	Mike Jordan 	
	Dean Pomerleau
	Steve Suddarth
=========================================================================
Connectionist Approaches to Symbol Grounding

Organizers: Georg Dorffner, Univ. Vienna; Michael Gasser, Indiana Univ.
            Stevan Harnad, Princeton Univ.

Abstract:
In recent years, there has been increasing discomfort with the
disembodied nature of symbols that is a hallmark of the symbolic
paradigm in cognitive science and artificial intelligence and at the
same time increasing interest in the potential offered by
connectionist models to ``ground'' symbols.
In ignoring the mechanisms by which their symbols get ``hooked up'' to
sensory and motor processes, that is, the mechanisms by which
intelligent systems develop categories, symbolists have missed out on
what is not only one of the more challenging areas in cognitive
science but, some would argue, the very heart of what cognition is about.
This workshop will focus on issues in neural network based
approaches to the grounding of symbols and symbol structures.
In particular, connectionist models of categorisation and
of label-category association will be discussed in the light of
the symbol grounding problem.

Invited presentations:
"Grounding Symbols in the Analog World of Objects: Can Neural
Nets Make the Connection?" Stevan Harnad, Princeton University

"Learning Perceptually Grounded Lexical Semantics"
Terry Regier, George Lakoff, Jerry Feldman, ICSI Berkeley

T.B.A.  Gary Cottrell, Univ. of California, San Diego

"Learning Perceptual Dimensions" Michael Gasser, Indiana University

"Symbols and External Embodiments - why Grounding has to Go
Two Ways" Georg Dorffner, University of Vienna

"Grounding Symbols on Conceptual Knowledge" Philippe Schyns, MIT
=========================================================================
Continuous Speech Recognition: Is there a connectionist advantage?

Organizer: Michael Franzini (maf at cs.cmu.edu)

Abstract:
This workshop will address the following questions: How do neural
networks  compare  to  the alternative technologies available for
speech recognition?  What evidence is available to  suggest  that
connectionism  may  lead  to  better  speech recognition systems?
What comparisons have been performed  between  connectionist  and
non-connectionist  systems,  and how ``fair'' are these comparis-
ons?  Which approaches to connectionist speech  recognition  have
produced  the  best  results, and which are likely to produce the
best results in the future?

Traditionally, the selection criteria for NIPS papers  reflect  a
much  greater  emphasis on theoretical importance of work than on
performance figures, despite the fact that  recognition  rate  is
one  of  the most important considerations for speech recognition
researchers (and often is  {\em the}  most  important  factor  in
determining  their  financial  support).   For  this reason, this
workshop -- to be oriented more towards performance  than  metho-
dology -- will be of interest to many NIPS participants.

The issue of connectionist vs. HMM performance in speech recogni-
tion  is  controversial in the speech recognition community.  The
validity of past comparisons is often disputed, as is the  funda-
mental  value  of  neural networks.  In this workshop, an attempt
will be made to address this issue and the questions stated above
by  citing  specific experimental results and by making arguments
with a theoretical basis.

Preliminary list of speakers:
	Ron Cole
	Uli Bodenhausen
	Hermann Hild
=========================================================================
Symbolic and Subsymbolic Information Processing in
	Biological Neural Circuits and Systems

Organizer: Vasant Honavar (honavar at iastate.edu)

Abstract:
Traditional information processing models in cognitive psychology
which became popular with the advent of the serial computer tended
to view cognition as discrete, sequential symbol processing.
Neural network or connectionist models offer an alternative paradigm
for modelling cognitive phenomena that relies on continuous, parallel
subsymbolic processing. Biological systems appear to combine both
discrete as well as continuous, sequential as well as parallel,
symbolic as well as subsymbolic information processing in various
forms at different levels of organization. The flow of neurotransmitter
molecules and of photons into receptors is quantal; the depolarization
and hyperpolarization of neuron membranes is analog; the genetic code
and the decoding processes appear to be digital; global interactions
mediated by neurotransmitters and slow waves appear to be both analog and
digital.

The purpose of this workshop is to bring together interested
computer scientists, neuroscientists, psychologists, mathematicians,
engineers, physicists and systems theorists to examine and discuss
specific examples as well as general principles (to the extent they can
be gleaned from our current state of knowledge) of information processing
at various levels of organization in biological neural systems.

The workshop will consist of several short presentations by participants
There will be ample time for informal presentations and discussion centering
around a number of key topics such as:

*  Computational aspects of symbolic v/s subsymbolic information processing
*  Coordination and control structures and processes in neural systems
*  Encoding and decoding structures and processes in neural systems
*  Generative structures and processes in neural systems
*  Suitability of particular paradigms for modelling specific phenomena
*  Software requirements for modelling biological neural systems

Invited presentations: TBA
Those interested in giving a presentation should write to honavar at iastate.edu
=========================================================================
Computational Issues in Neural Network Training

Organizers: Scott Markel and Roger Crane, Sarnoff Research

Abstract:
Many of the best practical neural network training results are report-
ed by researchers who use variants of back-propagation and/or develop
their own algorithms.  Few results are obtained by using classical nu-
merical optimization methods although such methods can be used effec-
tively for many practical applications.   Many competent researchers
have concluded, based on their own experience, that classical methods
have little value in solving real problems.  However, use of the best
commercially available implementations of such algorithms can help in
understanding numerical and computational issues that arise in all
training methods. Also, classical methods can be used effectively to
solve practical problems.  Examples of numerical issues that are ap-
propriate to discuss in this workshop include: convergence rates; lo-
cal minima; selection of starting points; conditioning (for higher
order methods); characterization of the error surface; ... .

Ample time will reserved for discussion and informal presentations. We
will encourage lively audience participation.
=========================================================================
Real Applications of Real Biological Circuits

Organizers: Richard Granger, UC Irvine and Jim Schwaber, Du Pont

Abstract:
The architectures, performance rules and learning rules of most artificial
neural networks are at odds with the anatomy and physiology of real
biological neural circuitry.  For example, mammalian telencephelon
(forebrain) is characterized by extremely sparse connectivity (~1-5%),
almost entirely lacks dense recurrent connections, and has extensive lateral
local circuit connections; inhibition is delayed-onset and relatively
long-lasting (100s of milliseconds) compared to rapid-onset brief excitation
(10s of milliseconds), and they are not interchangeable.  Excitatory
connections learn, but there is very little evidence for plasticity in
inhibitory connections.  Real synaptic plasticity rules are sensitive to
temporal information, are not Hebbian, and do not contain "supervision"
signals in any form related to those common in ANNs.

These discrepancies between natural and artificial NNs raise the question of
whether such biological details are largely extraneous to the behavioral and
computational utility of neural circuitry, or whether such properties may
yield novel rules that confer useful computational abilities to networks
that use them.  In this workshop we will explicitly analyze the power and
utility of a range of novel algorithms derived from detailed biology, and
illustrate specific industrial applicatons of these algorithms in the fields
of process control and signal processing.

It is anticipated that these issues will raise controversy, and half of
the workshop will be dedicated to open discussion.

Preliminary list of speakers:
	Jim Schwaber, DuPont
	Bbatunde Ogunnaike, DuPont
	Richard Granger, University of California, Irvine
	John Hopfield, Cal Tech
=========================================================================
Recognizing Unconstrained Handwritten Script

Organizers: Krishna Nathan, IBM and James A. Pittman, MCC

Abstract:
Neural networks have given new life to an old research topic, the
segmentation and recognition of on-line handwritten script.
Isolated handprinted character recognition systems are moving from
research to product development, and researchers have moved
forward to integrated segmentation and recognition projects.
However, the 'real world' problem is best described as one of
unconstrained handwriting recognition (often on-line) since it
includes both printed and cursive styles -- often within the same
word.

The workshop will provide a forum for participants to share ideas on
preprocessing, segmentation, and recognition techniques, and the use
of context to improve the performance of online handwriting recognition
systems. We will also discuss issues related to what constitutes
acceptable recognition performance.  The collection of training and
test data will also be addressed.
=========================================================================
Time Series Analysis and Predic....

Organizers: John Moody, Oregon Grad. Inst., Mike Mozer, Univ. of
            Colorado and Andreas Weigend, Xerox PARC

Abstract:
Several new techniques are now being applied to the problem of predicting
the future behavior of a temporal sequence and deducing properties of the
system that produced the time series. We will discuss both connectionist
and non-connectionist techniques.  Issues include algorithms and
architectures, model selection, performance measures, iterated vs long
term prediction, robust prediction and estimation, the number of degrees of
freedom of the system, how much noise is in the data, whether it is chaotic
or not, how the error grows with prediction time, detection and classification
of signals in noise, etc.  Half the available time has been reserved for
discussion and informal presentations. We will encourage lively audience
participation.

Invited presentations:
    Classical and Non-Neural Approaches: Advantages and Problems.
       (John Moody)
    Connectionist Approaches: Problems and Interpretations. (Mike Mozer)
    Beyond Prediction: What can we learn about the system? (Andreas Weigend)
    Physiological Time Series Modeling (Volker Tresp)
    Financial Forecasting (William Finnoff / Georg Zimmerman)
    FIR Networks (Eric Wan)
    Dimension Estimation (Fernando Pineda)
=========================================================================
Applications of VLSI Neural Networks

Organizer: Dave Andes, Naval Air Warfare Center

Abstract: This workshop will provide a forum for discussion
of the problems and opportunities for neural net hardware
systems which solve real problems under real time and space
constraints. Some of the most difficult requirements for
systems of this type come, not surprisingly, from the military.
Several examples of these problems and VLSI solutions will be
discussed in this working group. Examples from outside the
military will also be discussed. At least half the time will
be devoted to open discussion of the issues raised by the
experiences of those who have already applied VLSI based ANN
techniques to real world problems.

Preliminary list of speakers:
   Bill Camp, IBM Federal Systems
   Lynn Kern, Naval Air Warfare Center
   Chuck Glover, Oak Ridge National Lab
   Dave Andes, Naval Air Warfare Center

\enddata{text822, 0}