NIPS*2002 Workshops Abstracts

[Barak Pearlmutter]

   ****************************************************************

			 NIPS*2002 Workshops
	      December 12-14, 2002, Whistler BC, Canada
			  http://www.nips.cc

   ****************************************************************

				   
Workshop Schedule
=================


 The NIPS*2002 Workshops will be held at the Westin in Whistler BC,
 Canada, on Fri Dec 13 and Sat Dec 14, with sessions at 7:30-10:00am
 and 4:00-7:00pm.

    Two Day Workshops: Fri Dec 13 & Sat Dec 14

	Functional Neuroimaging
	Multi-Agent Learning
	Propagation on Cyclic Graphs

    One Day Workshops on Fri Dec 13

	Adaptation/Plasticity and Coding
	Bioinformatics
	Independent Component Analysis
	Neuromorphic Engineering
	Spectral Methods
	Statistics for Computational Experiments
	Unreal Data

    One Day Workshops on Sat Dec 13

	Learning Invariant Representations
	Learning Rankings
	Negative Results
	On Learning Kernels
	Quantum Neural Computing
	Thalamocortical Processing
	Universal Learning Algorithms


Workshop Descriptions
=====================


 TWO DAY WORKSHOPS (Friday & Saturday)


 Propagation Algorithms on Graphs with Cycles: Theory and Applications

	Shiro Ikeda, Kyushu Institute of Technology, Fukuoka, Japan
	Toshiyuki Tanaka, Tokyo Metropolitan University, Tokyo, Japan
	Max Welling, University of Toronto, Toronto, Canada

    Inference on graphs with cycles (loopy graphs) has drawn much
    attention in recent years. The problem arises in various fields
    such as AI, error-correcting codes, statistical physics, and image
    processing. Although exact inference is often intractable, much
    progress has been made in solving the problem approximately with
    local propagation algorithms. The aim of the workshop is to
    provide an overview of recent developments in methods related to
    belief propagation. We also encourage discussion of open
    theoretical problems and new possibilities for applications.


 Computational Neuroimaging: Foundations, Concepts & Methods

	Stephen J. Hanson, Rutgers University, Newark, NJ, USA
	Barak A. Pearlmutter, University of New Mexico, Albuquerque, NM, USA
	Stephen Strother, University of Minnesota, Minneapolis, MN, USA
	Lars Kai Hansen, Technical University of Denmark, Lyngby, Denmark
	Benjamin Martin Bly, Rutgers University, Newark, NJ, USA

    This workshop will concentrate on the foundations of neuroimaging,
    including the relation between neural firing and BOLD, fast fMRI,
    and diffusion methods.  The first day includes speakers on new
    Methods for Multivariate analysis using fMRI especially as they
    relate to Neural Modeling (ICA, SVM, or other ML methods), which
    will slip into the next morning, with cognitive neuroscience talks
    involving Network and specific Neural Modeling approaches to
    cognitive function on day two.


 Multi-Agent Learning: Theory and Practice

	Gerald Tesauro, IBM Research, NY, USA
	Michael L. Littman, Rutgers University, New Brunswick, NJ, USA

    Machine learning in a multi-agent system, where learning agents
    interact with other agents that are also simultaneously learning,
    poses a radically different set of issues from those arising in
    normal single-agent learning in a stationary environment.  This
    topic is poorly understood theoretically but seems ripe for
    progress by building upon many recent advances in RL and in
    Bayesian, game-theoretic, decision-theoretic, and evolutionary
    learning.  At the same time, learning is increasingly vital in
    fielded applications of multi-agent systems.  Many application
    domains are envisioned in which teams of software agents or robots
    learn to cooperate to achieve global objectives.  Learning may
    also be essential in many non-cooperative domains such as
    economics and finance, where classical game-theoretic solutions
    are either infeasible or inappropriate.  This workshop brings
    together researchers studying multi-agent learning from a variety
    of perspectives.  Our invited speakers include leading AI
    theorists, applications developers in fields such as robotics and
    e-commerce, as well as social scientists studying learning in
    multi-player human-subject experiments. Slots are also available
    for contributed talks and/or posters.


 ONE DAY WORKSHOPS (Friday)


 The Role of Adaptation/Plasticity in Neuronal Coding

	Garrett B. Stanley, Harvard University, Cambridge, MA, USA
	Tai Sing Lee, Carnegie Mellon University, Pittsburgh, PA, USA

    A ubiquitous characteristic of neuronal processing is the ability
    to adapt to an ever changing environment on a variety of different
    time scales.  Although the different forms of adaptation/
    plasticity have been studied for some time, their role in the
    encoding process is still not well understood. The most widely
    utilized measures assume time-invariant encoding dynamics even
    though mechanisms serving to modify coding properties are
    continually active in all but the most artificial laboratory
    conditions.  Important questions include: (1) how do encoding
    dynamics and/or receptive field properties change with time and
    the statistics of the environment?, (2) what are the underlying
    sources of these changes?, (3) what are the resulting effects on
    information transmission and processing in the pathway?, and (4)
    can the mechanisms of plasticity/adaptation be understood from a
    behavioral perspective?  It is the goal of this workshop to
    discuss neuronal coding within several different experimental
    paradigms, in order to explore these issues that have only
    recently been addressed in the literature.


 Independent Component Analysis and Beyond

	Stefan Harmeling, Fraunhofer FIRST, Berlin, Germany
	Luis Borges de Almeida, INESC ID, Lisbon, Portugal
	Erkki Oja, HUT, Helsinki, Finland
	Dinh-Tuan Pham, LMC-IMAG, Grenoble, France

    Independent component analysis (ICA) aims at extracting unknown
    hidden factors/components from multivariate data using only the
    assumption that the unknown factors are mutually independent.
    Since the introduction of ICA concepts in the early 80s in the
    context of neural networks and array signal processing, many new
    successful algorithms have been proposed that are now
    well-established methods. Since then, diverse applications in
    telecommunications, biomedical data analysis, feature extraction,
    speech separation, time-series analysis and data mining have been
    reported.  Notably of special interest for the NIPS community are,
    first, the application of ICA techniques to process multivariate
    data from various neuro-physiological recordings and second, the
    interesting conceptual parallels to information processing in the
    brain.  Recently exciting developments have moved the field
    towards more general nonlinear or nonindependent source separation
    paradigms. The goal of the planed workshop is to bring together
    researchers from the different fields of signal processing,
    machine learning, statistics and applications to explore these new
    directions.


 Spectral Methods in Dimensionality Reduction, Clustering, and
 Classification

	Josh Tenenbaum, M.I.T., Cambridge, MA, USA
	Sam Roweis, University of Toronto, Ontario, Canada

    Data-driven learning by local or greedy parameter update
    algorithms is often a painfully slow process fraught with local
    minima.  However, by formulating a learning task as an appropriate
    algebraic problem, globally optimal solutions may be computed
    efficiently in closed form via an eigendecomposition.
    Traditionally, this spectral approach was thought to be applicable
    only to learning problems with an essentially linear structure,
    such as principal component analysis or linear discriminant
    analysis. Recently, researchers in machine learning, statistics,
    and theoretical computer science have figured out how to cast a
    number of important nonlinear learning problems in terms amenable
    to spectral methods. These problems include nonlinear
    dimensionality reduction, nonparameteric clustering, and nonlinear
    classification with fully or partially labeled data.  Spectral
    approaches to these problems offer the potential for dramatic
    improvements in efficiency, accuracy, optimality and
    reproducibility relative to traditional iterative or greedy
    learning algorithms.  Furthermore, numerical methods for spectral
    computations are extremely mature and well understood, allowing
    learning algorithms to benefit from a long history of
    implementation efficiencies in other fields. The goal of this
    workshop is to bring together researchers working on spectral
    approaches across this broad range of problem areas, for a series
    of talks on state-of-the-art research and discussions of common
    themes and open questions.


 Neuromorphic Engineering in the Commercial World

	Timothy Horiuchi, University of Maryland, College Park, MD, USA
	Giacomo Indiveri, University-ETH Zurich, Zurich, Switzerland
	Ralph Etienne-Cummings, University of Maryland, College Park, MD, USA

    We propose a one-day workshop to discuss strategies, opportunities
    and success stories in the commercialization of neuromorphic
    systems. Towards this end, we will be inviting speakers from
    industry and universities with relevant experience in the
    field. The discussion will cover a broad range of topics, from
    visual and auditory processing to olfaction and locomotion,
    focusing specifically on the key elements and ideas for
    successfully transitioning from neuroscience to commercialization.


 Statistical Methods for Computational Experiments in Visual Processing
 and Computer Vision

	Ross Beveridge, Colorado State University, Colorado, USA
	Bruce Draper, Colorado State University, Colorado, USA
	Geof Givens, Colorado State University, Colorado, USA
	Ross J. Micheals, NIST, Maryland, USA
	Jonathon Phillips, DARPA & NIST, Maryland, USA

    In visual processing and computer vision, computational
    experiments play a critical role in explaining algorithm and
    system behavior.  Disciplines such as psychophysics and medicine
    have a long history of designing experiments.  Vision researchers
    are still learning how to use computational experiments to explain
    how systems behave in complex domains.  This workshop will focus
    on new and better experiment experimental methods in the context
    of visual processing and computer vision.


 Unreal Data: Principles of Modeling Nonvectorial Data

	Alexander J. Smola, Australian National Univ., Canberra, Australia
	Gunnar Raetsch, Australian National Univ., Canberra, Australia
	Zoubin Ghahramani, University College London, London, UK

    A large amount of research in machine learning is concerned with
    classification and regression for real-valued data which can
    easily be embedded into a Euclidean vector space. This is in stark
    contrast with many real world problems, where the data is often a
    highly structured combination of features, a sequence of symbols,
    a mixture of different modalities, may have missing variables,
    etc. To address the problem of learning from non-vectorial data,
    various methods have been proposed, such as embedding the
    structures in some metric spaces, the extraction and selection of
    features, proximity based approaches, parameter constraints in
    Graphical Models, Inductive Logic Programming, Decision Trees,
    etc. The goal of this workshop is twofold. Firstly, we hope to
    make the machine learning community aware of the problems arising
    from domains where non-vectorspace data abounds and to uncover the
    pitfalls of mapping such data into vector spaces. Secondly, we
    will try to find a more uniform structure governing methods for
    dealing with non-vectorial data and to understand what, if any,
    are the principles underlying the modeling of non-vectorial data.


 Machine Learning Techniques for Bioinformatics

	Colin Campbell, University of Bristol, UK
	Phil Long, Genome Institute of Singapore

    This workshop will cover the development and application of
    machine learning techniques in application to molecular biology.
    Contributed papers are welcome from any topic relevant to this
    theme including, but not limited to, analysis of expression data,
    promoter analysis, protein structure prediction, protein homology
    detection, detection of splice junctions, and phylogeny, for
    example.  Contributions are most welcome which propose new
    algorithms or methods, rather than the use of existing techniques.
    In addition to contributed papers we expect to have several
    tutorials covering different areas where machine learning
    techniques are have been successfully applied in this domain.


 ONE DAY WORKSHOPS (Saturday)


 Thalamocortical Processing in Audition and Vision

	Tony Zador, Cold Spring Harbor Lab., Cold Spring Harbor, NY, USA
	Shihab Shamma, University of Maryland, College Park, MD, USA

    All sensory information (except olfactory) passes through the
    thalamus before reaching the cortex. Are the principles governing
    this thalamocortical transformation shared across sensory
    modalities? This workshop will investigate this question in the
    context of audition and vision. Questions include: Do the LGN and
    MGN play analogous roles in the two sensory modalities? Are the
    cortical representations of sound and light analogous?
    Specifically, the idea is to talk about cortical processing (as
    opposed to purely thalamic), how receptive fields are put together
    in the cortex, and the implications of these ideas to the nature
    of information being encoded and extracted at the cortex.


 Learning of Invariant Representations

	Konrad Paul Koerding, ETH/UNI Zuerich, Switzerland
	Bruno. A. Olshausen, U.C. Davis & RNI, CA, USA

    Much work in recent years has shown that the sensory coding
    strategies employed in the nervous systems of many animals is well
    matched to the statistics of their natural environment.  For
    example, it has been shown that lateral inhibition occuring in the
    retina may be understood in terms of a decorrelation or
    `whitening' strategy (Srinivasan et al., 1982; Atick & Redlich,
    1992), and that the receptive properties of cortical neurons may
    be understood in terms of sparse coding or ICA (Olshausen & Field,
    1996; Bell & Sejnowski, 1997; van Hateren & van der Schaaf,
    1998). However, most of these models do not address the question
    of which properties of the environment are interesting or relevant
    and which others are behaviourally insignificant. The purpose of
    this workshop is to focus on unsupervised learning models that
    attempt to represent features of the environment which are
    invariant or insensitive to variations such as position, size, or
    other factors.


 Quantum Neural Computing

	Elizabeth C. Behrman, Wichita State University, Wichita, KS, USA
	James E. Steck, Wichita State University, Wichita, KS, USA

    Recently there has been a resurgence of interest in quantum
    computers because of their potential for being very much smaller
    and very much faster than classical computers, and because of
    their ability in principle to do hereofore impossible
    calculations, such as factorization of large numbers in polynomial
    time.  We will explore ways to implement biologically inspired
    quantum computing in network topologies, thus exploiting both the
    intrinsic advantages of quantum computing and the adaptability of
    neural computing. This workshop will follow up on our very
    successful NIPS 2000 workshop and the IJCNN 2001 Special Session.
    Aspects/approaches to be explored will include: quantum hardware,
    e.g., SQUIDs, nmr, trapped ions, quantum dots, and molecular
    computing; theoretical and practical limits to quantum and quantum
    neural computing, e.g. noise, error correction, and decoherence;
    and simulations.


 Universal Learning Algorithms and Optimal Search

	Juergen Schmidhuber, IDSIA, Manno-Lugano, Switzerland
	Marcus Hutter, IDSIA, Manno-Lugano, Switzerland

    Recent theoretical and practical advances are currently driving a
    renaissance in the fields of Universal Learners (rooted in
    Solomonoff's universal induction scheme, 1964) and Optimal Search
    (rooted in Levin's universal search algorithm, 1973). Both are
    closely related to the theory of Kolmogorov complexity. The new
    millennium has brought several significant developments including:
    Sharp expected loss bounds for universal sequence predictors,
    theoretically optimal reinforcement learners for general
    computable environments, computable optimal predictions based on
    natural priors that take algorithm runtime into account, and
    practical, bias-optimal, incremental, universal search algorithms.
    Topics will also include: Practical but general MML/MDL/SRM
    approaches with theoretical foundation, weighted majority
    approaches, and no free lunch theorems.


 On Learning Kernels

	Nello Cristianini, U.C. Davis, California, USA
	Tommi Jaakkola, M.I.T., Massachusetts, USA
	Michael I. Jordan, U.C. Berkeley, California, USA
	Gert R.G. Lanckriet, U.C. Berkeley, California, USA

    Recent theoretical advances and experimental results have drawn
    considerable attention to the use of kernel methods in learning
    systems. For the past five years, a growing community has been
    meeting at the NIPS workshops to discuss the latest progress in
    learning with kernels. Recent research in this area addresses the
    problem of learning the kernel itself from data.  This subfield is
    becoming an active research area, offering a challenging interplay
    between statistics, advanced convex optimization and information
    geometry. It presents a number of interesting open problems. The
    workshop has two goals.  First, it aims at discussing
    state-of-the-art research on 'learning the kernel', as well as
    giving an introduction to some of the new techniques used in this
    subfield. Second, it offers a meeting point for a diverse
    community of researchers working on kernel methods.  As such,
    contributions from ALL subfields in kernel methods are welcome and
    will be considered for a poster presentation, with priority to
    very recent results. Furthermore, contributions on the main theme
    of learning kernels will be considered for oral presentations.
    Deadline for submissions: Nov 15, 2002.


 Negative Results and Open Problems

	Isabelle Guyon, Clopinet, California, USA

    In mathematics and theoretical computer science, exhibiting
    counter examples is part of the established scientific method to
    rule out wrong hypotheses. Yet, negative results and counter
    examples are seldom reported in experimental papers, although they
    can be very valuable.  Our workshop will be a forum to freely
    discuss negative results and introduce the community to
    challenging open problems. This may include reporting experimental
    results of principled algorithms that obtain poor performance
    compared to seemingly dumb heuristics; experimental results that
    falsify an existing theory; counter examples to a generally
    admitted conjecture; failure to find a solution to a given problem
    after various attempts; and failure to demonstrate the advantage
    of a given method after various attempts. If you have interesting
    negative results (not inconclusive results) or challenging open
    problems, you may submit an abstract before November 15, 2002.


 Beyond Classification and Regression: Learning Rankings, Preferences,
 Equality Predicates, and Other Structures

	Rich Caruana, Cornell University, NY, USA
	Thorsten Joachims, Cornell University, NY, USA

    Not all supervised learning problems fit the classification/
    regression function-learning model. Some problems require
    predictions other than values or classes. For example, sometimes
    the magnitude of the values predicted for cases are not important,
    but the ordering these values induce is important. This workshop
    addresses supervised learning problems where either the goal of
    learning or the input to the learner is more complex than in
    classification and regression. Examples of such problems include
    learning partial or total orderings, learning equality or match
    rules, learning to optimize non-standard criteria such as
    Precision and Recall or ROC Area, using relative preferences as
    training examples, learning graphs and other structures, and
    problems that benefit from these approaches (e.g., text retrieval,
    medical decision making, protein matching). The goal of this
    one-day workshop is to discuss the current state-of-the-art, and
    to inspire research on new algorithms and problems. To submit an
    abstract, see http://www.cs.cornell.edu/People/tj/ranklearn.



More extensive information is available on the NIPS web page
http://www.nips.cc, which has links to the pages maintained by each
individual workshop.

The number of workshop proposals was particularly high this year.  All
together there will be seventeen NIPS*2002 workshops, of which three
will last for two days, for a total of twenty workshop-days: a new
record.  We anticipate a great year not just in the number of
workshops and in their quality, but in attendance as well: projections
indicate that the workshops may surpass the main conference in total
number of participants.