[UAI] NIPS*2001 workshops

From: Richard Zemel (zemel@cs.toronto.edu)
Date: Thu Oct 11 2001 - 08:31:37 PDT

  • Next message: Richard Zemel: "[UAI] NIPS*2001 registration"

              * * * Post-NIPS*2001 Workshops * * *
              * * * Whistler, BC, CANADA * * *
              * * * December 7-8, 2001 * * *

    The NIPS*2001 Workshops will be on Friday and Saturday, December 7/8,
    in Whistler, BC, Canada, following the main NIPS conference in
    Vancouver Monday-Thursday, December 3-6.

    This year there are 19 workshops:

      Activity-Dependent Synaptic Plasticity
      Artificial Neural Networks in Safety-Related Areas
      Brain-Computer Interfaces
      Causal Learning and Inference in Humans & Machines
      Competition: Unlabeled Data for Supervised Learning
      Computational Neuropsychology
      Geometric Methods in Learning
      Information & Statistical Structure in Spike Trains
      Kernel-Based Learning
      Knowledge Representation in Meta-Learning
      Machine Learning in Bioinformatics
      Machine Learning Methods for Images and Text
      Minimum Description Length
      Multi-sensory Perception & Learning
      Neuroimaging: Tools, Methods & Modeling
      Occam's Razor & Parsimony in Learning
      Preference Elicitation
      Quantum Neural Computing
      Variable & Feature Selection

    Some workshops span both days, while others will be only one day long.
    One-day workshops will be assigned to friday or saturday by October 14.
    Please check the web page after this time for individual dates.

    All workshops are open to all registered attendees. Many workshops
    also invite submissions. Submissions, and questions about individual
    workshops, should be directed to the individual workshop organizers.
    Included below is a short description of most of the workshops.
    Additional information (including web pages for the individual
    workshops) is available at the NIPS*2001 Web page:

       http://www.cs.cmu.edu/Groups/NIPS/

    Information about registration, travel, and accommodations for the
    main conference and the workshops is also available there.

    Whistler is a ski resort a few hours drive from Vancouver. The daily
    workshop schedule is designed to allow participants to ski half days,
    or enjoy other extra-curricular activities. Some may wish to extend
    their visit to take advantage of the relatively low pre-season rates.

    We look forward to seeing you in Whistler.

    Virginia de Sa and Barak Pearlmutter
       NIPS Workshops Co-chairs

    - -------------------------------------------------------------------------

    Activity-dependent Synaptic Plasticity

            Paul Munro, Larry Abbott
            http://www.pitt.edu/~pwm/plasticity

        While the mathematical and cognitive aspects of rate-based
        Hebb-like rules have been broadly explored, relatively little is
        known about the possible role of STDP at the computational
        level. Hebbian learning in neural networks requires both
        correlation-based synaptic plasticity and a mechanism that induces
        competition between different synapses. Spike-timing-dependent
        synaptic plasticity is especially interesting because it combines
        both of these elements in a single synaptic modification
        rule. Some recent work has examined the possibility that STDP may
        underlie older models, such as Hopfield networks or the BCM
        rule. Temporally dependent synaptic plasticity is attracting a
        rapidly growing amount of attention in the computational
        neuroscience community. The change in synaptic efficacy arising
        from this form of plasticity is highly sensitive to temporal
        correlations between different presynaptic spike
        trains. Furthermore, it can generate asymmetric and directionally
        selective receptive fields, a result supported by experiments on
        experience-dependent modifications of hippocampal place
        fields. Finally, spike-timing-dependent plasticity automatically
        balances excitation and inhibition producing a state in which
        neuronal responses are rapid but highly variable. The major goals
        of the workshop are:

        1. To review current experimental results on
           spike-timing-dependent synaptic plasticity and related effects.

        2. To discuss models and mechanisms for this form of synaptic plasticity.

        3. To explore the relationship of STDP with other approaches.

        4. To reconcile the rate-based and spike-based plasticity data
           with a unified theoretical framework (very optimistic!)..

    - -------------------------------------------------------------------------

    Artificial Neural Networks in Safety-Related Areas:
    Applications and Methods for Validation and Certification

            J. Schumann, P. Lisboa, R. Knaus
            http://ase.arc.nasa.gov/people/schumann/workshops/NIPS2001

        Over the recent years, Artificial Neural Networks have found their
        way into various safety-related and safety-critical areas, for
        example, power generation and transmission, transportation,
        avionics, environmental monitoring and control, medical
        applications, and consumer products. Applications range from
        classification to monitoring and control. Quite often, these
        applications proved to be highly successful, leading from pure
        research prototypes into serious experimental systems (e.g., a
        neural-network-based flight-control system test-flown on a NASA
        F-15ACTIVE) or commercial products (e.g., Sharp's
        Logi-cook). However, the general question of how to make sure that
        the ANN-based system performs as expected in all cases has not yet
        been addressed satisfactorily. All safety-related software
        applications require careful verification and validation (V&V) of
        the software components, ranging from extended testing to
        full-fledged certification procedures. However, for neural-network
        based systems, a number of specific issues have to be
        addressed. For example, a lack of a concise plant model, often a
        major reason to use a ANN in the first place, makes traditional
        approaches to V&V impossible.

        In this workshop, we will address such issues. In particular, we
        will discuss the following (non-exhaustive list of) topics: *
        theoretical methodologies to characterise the properties of ANN
        solutions, e.g., multiple realisations of a particular network and
        ways of managing this * fundamental software approaches to V&V and
        implications for ANNs, e.g., the application of FMEA * statistical
        (Bayesian) methods and symbolic techniques like rule extraction
        with subsequent V&V to assess and guarantee the performance of a
        ANN * dynamic monitoring of the ANN's behavior * stability proofs
        for control of dynamical systems with ANNs * principled approaches
        to design assurance, risk assessment, and performance evaluation
        of systems with ANNs * experience of application and certification
        of ANNs for safety-related applications * V&V techniques suitable
        for on-line trained and adaptive systems

        This workshop aims to bring together researchers who have applied
        ANNs in safety-related areas and actually addressed questions of
        demonstrating flawless operation of the ANN, researchers working
        on theoretical topics of convergence and performance assessment,
        researchers in the area of nonlinear adaptive control, and
        researchers from the area of formal methods in software design for
        safety-critical systems. Many prototypical/experimental
        application of neural networks in safety-related areas have
        demonstrated their usefulness successfully. But ANN applicability
        in safety-critical areas is substantially limited because of a
        lack of methods and techniques for verification and
        validation. Currently, there is no silver bullet for V&V in
        traditional software, and with the more complicated situation for
        ANNs, none is expected here in the short run. However, any result
        can have substantial impact in this field.

    - -------------------------------------------------------------------------

    Brain-Computer Interfaces

            Lucas Parra, Paul Sajda, Klaus-Robert Mueller
            http://newton.bme.columbia.edu/bci

    - -------------------------------------------------------------------------

    Causal learning and inference in humans and machines

            T. Griffiths, J. Tenenbaum, T. Kushnir, K. Murphy, A. Gopnik
            http://www-psych.stanford.edu/~jbt/causal-workshop.html

        The topic of causality has recently leapt to the forefront of
        theorizing in the fields of cognitive science, statistics, and
        artificial intelligence. The main objective of this workshop is to
        explore the potential connections between research on causality in
        the these three fields. There has already been much productive
        cross-fertilization: the development of causal Bayes nets in the
        AI community has often had a strong psychological motivation, and
        recent work by several groups in cognitive science has shown that
        some elementary but important aspects of how people learn and
        reason about causes may be best explained by theories based on
        causal Bayes nets.  Yet the most important questions lay wide
        open. Some examples of the questions we hope to address in this
        workshop include:

        * Can we scale up Bayes-net models of human causal learning and
        inference from microdomains with one or two causes and effects to
        more realistic large-scale domains?

        * What would constitute strong empirical tests of large-scale
        Bayes net models of human causal reasoning?

        * Do approximation methods for inference and learning on large
        Bayes nets have anything to do with human cognitive processes?

        * What are the relative roles of passive observation and active
        manipulation in causal learning?

        * What is the relation between psychological and computational
        notions of causal independence?

        The workshop will last one day. Most of the talks will be
        invited, but we welcome contributions for short talks by
        researchers in AI, statistics or cognitive science would like to
        make connections between these fields. Please contact one of the
        organizers if you are interested in participating. For more
        information contact Josh Tenenbaum (jbt@psych.stanford.edu) or
        Alison Gopnik (gopnik@socrates.berkeley.edu).

    - -------------------------------------------------------------------------

    Competition: Unlabeled Data for Supervised Learning

            Stefan C. Kremer, Deborah A. Stacey
            http://q.cis.uoguelph.ca/~skremer/NIPS2001/

        Recently, there has been much interest in applying techniques that
        incorporate knowledge from unlabeled data into systems performing
        supervised learning. The potential advantages of such techniques
        are obvious in domains where labeled data is expensive and
        unlabeled data is cheap. Many such techniques have been proposed,
        but only recently has any effort been made to compare the
        effectiveness of different approaches on real world problems.

        This web-site presents a challenge to the proponents of methods to
        incorporate unlabeled data into supervised learning. Can you
        really use unlabeled data to help train a supervised
        classification (or regression) system? Do recent (and not so
        recent) theories stand up to the data test?

        On this web-site you can find challenge problems where you can try
        out your methods head-to-head against anyone brave enough to face
        you. Then, at the end of the contest we will release the results
        and find out who really knows something about using unlabeled
        data, and if unlabeled data are really useful or we are all just
        wasting our time. So ask yourself, are you (and your theory) up to
        the challenge?? Feeling lucky???

    - -------------------------------------------------------------------------

    Computational Neuropsychology

            Sara Solla, Michael Mozer, Martha Farah
            http://www.cs.colorado.edu/~mozer/nips2001workshop.html

        The 1980's saw two important developments in the sciences of the
        mind: The development of neural network models in cognitive
        psychology, and the rise of cognitive neuroscience. In the 1990's,
        these two separate approaches converged, and one of the results
        was a new field that we call "Computational Neuropsychology." In
        contrast to traditional cognitive neuropsychology, computational
        neuropsychology uses the concepts and methods of computational
        modeling to infer the normal cognitive architecture from the
        behavior of brain-damaged patients. In contrast to traditional
        neural network modeling in psychology, computational
        neuropsychology derives constraints on network architectures and
        dynamics from functional neuroanatomy and neurophysiology.
        Unfortunately, work in computational neuropsychology has had
        relatively little contact with the Neural Information Processing
        Systems (NIPS) community. Our workshop aims to expose the NIPS
        community to the unusual patient cases in neuropsychology and the
        sorts of inferences that can be drawn from these patients based on
        computational models, and to expose researchers in computational
        neuropsychology to some of the more sophisticated modeling
        techniques and concepts that have emerged from the NIPS community
        in recent years.

        We are interested in speakers from all aspects of neuropsychology,
        including:

            * attention (neglect)
            * visual and auditory perception (agnosia)
            * reading (acquired dyslexia)
            * face recognition (prosopagnosia)
            * memory (Alzheimer's, amnesia, category-specific deficits)
            * language (aphasia)
            * executive function (schizophrenia, frontal deficits).

        Contact Sara Solla (solla@nwu.edu) or Mike Mozer
        (mozer@colorado.edu) if you are interested in speaking at the
        workshop.

    - -------------------------------------------------------------------------

    Geometric Methods in Learning workshop

            Amir Assadi
            http://www.lmcg.wisc.edu/bioCVG/events/NIPS2001/NIPS2001Wkshp.htm
            http://www.lmcg.wisc.edu/bioCVG

        The purpose of this workshop is to attract the attention of the
        learning community to geometric methods and to take on an
        endeavor:

        1. To lay out a geometric paradigm for formulating profound ideas
           in learning;

        2. To facilitate the development of geometric methods suitable of
           investigation of new ideas in learning theory.

        Today's continuing advances in computation make it possible to
        infuse geometric ideas into learning that otherwise would have
        been computationally prohibitive. Nonlinear dynamics in brain-like
        complex systems has created great excitement, offering a broad
        spectrum of new ideas for discovery of parallel-distributed
        algorithms, a hallmark of learning theory. By having great
        overlap, geometry and nonlinear dynamics together offer a
        complementary and more profound picture of the physical world and
        how it interacts with the brain, the ultimate learning system.

        Among the discussion topics, we envision the following:
        information geometry, differential topological methods for turning
        local estimates into global quantities and invariants, Riemannian
        geometry and Feynman path integration as a framework to explore
        nonlinearity, advanced in complex dynamical system theory in the
        context of learning and dynamic information processing in brain,
        and information theory of massive data sets. As before, in our
        discussion sessions we will also examine the potential impact of
        learning theory on future development of geometry, and report on
        new examples of new vistas on the impact of learning theoretic
        parallel-distributed algorithms on research in mathematics.

        With 3 years of meetings, we are in a position to plan a volume
        based on the materials for the workshops and other contributions
        to be proposed to the NIPS Program Committee.

    - -------------------------------------------------------------------------

    Information and Statistical Structure in Spike Trains

            Jonathon D. Victor
            http://www-users.med.cornell.edu/~jdvicto/nips2001.html

        Understanding how neurons represent and manipulate information in
        their spike trains is one of the major fundamental problems in
        neuroscience. Moreover, advances towards its solution will rely
        on a combination of appropriate theoretical, computational, and
        experimental strategies. Meaningful and reliable statistical
        analyses, including calculation of information and related
        quantities, are at the basis of understanding neural information
        processing. The accuracy and precision of statistical analyses and
        empirical information estimates depend strongly on the amount and
        quality of the data available, and on the assumptions that are
        made in order to apply the formalisms to a laboratory data
        set. These assumptions typically relate to the neural transduction
        itself (e.g., linearity or stationarity) and to the statistics of
        the spike trains (e.g., correlation structure). There are numerous
        approaches to conducting statistical analyses and estimating
        information-theoretic quantities, and there are also some major
        differences in findings across preparations. It is unclear to what
        extent these differences represent fundamental biological
        differences, differences in what is being measured, or
        methodological biases. Specific areas of focus will include:
        Theoretical and experimental approaches to analyze multineuronal
        spiking activity; Bursting, rhythms, and other endogenous
        patterns; Is "Poisson-like" a reasonable approximation to spike
        train stochastic structure?; How do we formulate alternative
        models to Poisson?; How do we evaluate model goodness-of-fit?

        A limited number of slots are available for contributed
        presentations. Individuals interested in presenting a talk
        (approximately 20 minutes, with 10 to 20 minutes for discussion)
        should submit a title and abstract, 200-300 words, to the
        organizers, Jonathan D. Victor (jdvicto@med.cornell.edu) and Emery
        Brown (brown@neurostat.mgh.harvard.edu) by October 12, 2001.

    - -------------------------------------------------------------------------

    Workshop on New Directions in Kernel-Based Learning Methods

            Chris Williams, Craig Saunders, Matthias Seeger, John Shawe-Taylor
            http://www.cs.rhul.ac.uk/colt/nipskernel.html

        The aim of the workshop is to present new perspectives and new
        directions in kernel methods for machine learning. Recent
        theoretical advances and experimental results have drawn
        considerable attention to the use of kernel functions in learning
        systems. Support Vector Machines, Gaussian Processes, kernel PCA,
        kernel Gram-Schmidt, Bayes Point Machines, Relevance and Leverage
        Vector Machines, are just some of the algorithms that make crucial
        use of kernels for problems of classification, regression, density
        estimation, novelty detection and clustering. At the same time as
        these algorithms have been under development, novel techniques
        specifically designed for kernel-based systems have resulted in
        methods for assessing generalisation, implementing model
        selection, and analysing performance. The choice of model may be
        simply determined by parameters of the kernel, as for example the
        width of a Gaussian kernel. More recently, however, methods for
        designing and combining kernels have created a toolkit of options
        for choosing a kernel in a particular application. These methods
        have extended the applicability of the techniques beyond the
        natural Euclidean spaces to more general discrete structures.

        The workshop will provide a forum for discussing results and
        problems in any of the above mentioned areas. But more
        importantly, by the structure of the workshop we hope to examine
        the future directions and new perpsectives that will keep the
        field lively and growing.

        We seek two types of contributions:

         1) Contributed 20 minutes talks that offer new directions
            (serving as a focal point for the general discussions)

         2) Posters of new ongoing work, with associated spotlight
            presentations (summarising current work and serving as a
            springboard for individual discussion).

        Important Dates:

         Submission of extended abstracts: 15th October 2001.
         Notification of acceptance: Early November.

         Submission Procedure: Extended abstracts in .ps or .pdf formats
         (only) should be e-mailed to nips-kernel-workshop@cs.rhul.ac.uk

    - -------------------------------------------------------------------------

    Knowledge Representation In Meta-Learning

            Ricardo Vilalta
            http:www/research.ibm.com/MetaLearning

        Learning across multiple related tasks, or improving learning
        performance over time, requires knowledge be transferred across
        tasks. In many classification algorithms, successive applications
        of the algorithm over the same data always produces the same
        hypothesis; no knowledge is extracted across tasks. Knowledge
        across tasks can be used to construct meta-learners able to
        improve the quality of the inductive bias through experience. To
        attain this goal, different pieces of knowledge are needed. For
        example, how can we characterize those tasks that are most
        favorable to a particular classification algorithm? On the other
        hand, What forms of bias are most favorable for certain tasks? Are
        there invariant transformations inherent to a domain that can be
        captured when learning across tasks? The goal of the workshop is
        to discuss alternative ways of knowledge representation in
        meta-learning with the idea of achieving new forms of bias
        adaptation.

        Important Dates: Paper submission: Nov 1, 2001. Notification of
        acceptance: Nov 12, 2001. Camera-ready copy: Nov 26, 2001.

    - -------------------------------------------------------------------------

    Machine Learning Techniques for Bioinformatics

            Colin Campbell, Shayan Mukherjee
            http://lara.enm.bris.ac.uk/cig/nips01/nips01.htm

        There has been significant recent interest in the development of
        new methods for functional interpretation of gene expression data
        derived from cDNA microarrays and related technologies. Analysis
        frequently involves classification, regression, feature selection,
        outlier detection and cluster analysis, for example. To provide a
        focus, this topic be the main theme for this one-day Workshop,
        though contributions in related areas of bioinformatics are
        welcome. Contributed papers should ideally be in the area of new
        algorithmic or theoretical approaches to analysing such datasets
        as well as biologically interesting applications and validation of
        existing algorithms. To make sure the Workshop relates to issues
        of real importance to experimentalists there will be four invited
        tutorial talks to introduce microarray technology, illustrate
        particular case studies and discuss issues relevant to eventual
        clinical application. The invited speakers are Pablo Tamayo or
        Todd Golub (Whitehead Institute, MIT), Dan Notterman (Princeton
        University), Roger Bumgarner (University of Washington) and
        Richard Simon (National Cancer Institute). The invited speakers
        have been involved in the preparation of well-known datasets and
        studies of expression analysis for a variety of cancers. Authors
        wishing to contribute papers should submit a title and extended
        abstract to both organisers (C.Campbell@bris.ac.uk and
        sayan@mit.edu) before 14th October 2001. Further details about
        this workshop and the final schedule are available from the
        workshop webpage.

    - -------------------------------------------------------------------------

    Machine Learning Methods for Images and Text

            Thomas Hofmann, Jaz Kandola, Tomaso Poggio, John Shawe-Taylor
            http://www.cs.rhul.ac.uk/colt/nipstext.html

        The aim of the workshop is to present new perspectives and new
        directions in information extraction from structured and
        semi-structured data for machine learning. The goal of this
        workshop is to investigate extensions of modern statistical
        learning techniques for applications in the domains of
        categorization and retrieval of information for example text,
        video and sound, as well as to their combination --
        multimedia. The focus will be on exploring innovative and
        potentially groundbreaking machine learning technologies as well
        as on identifying key challenges in information access, such as
        multi-class classification, partially labeled examples and the
        combination of evidence from separate multimedia domains. The
        workshop aims to bring together an interdisciplinary group of
        international researchers from machine learning, information
        retrieval, computational linguistics, human-computer interaction,
        and digital libraries for discussing results and dissemination of
        ideas, with the objective of highlighting new research
        directions. The workshop will provide a forum for discussing
        results and problems in any of the above mentioned areas. But more
        importantly, by the structure of the workshop we hope to examine
        the future directions and new perpsectives that will keep the
        field lively and growing. We seek two types of contributions:

        1) Contributed 20 minutes talks that offer new directions (serving
           as a focal point for the general discussions)

        2) Posters of new ongoing work, with associated spotlight
           presentations (summarising current work and serving as a
           springboard for individual discussion).

        Important Dates: Submission of extended abstracts: 15th October
        2001. Notification of acceptance: 2nd November 2001.

        Submission Procedure: Extended abstracts in .ps or .pdf formats
        (only) should be e-mailed to nips-text-workshop@cs.rhul.ac.uk by
        15th October 2001. Extended abstracts should be 2-4 sides of A4.
        The higlighting of a confernce-style group for the paper is not
        necessary, however the indication of a group and/or keywords would
        be helpful.

    - -------------------------------------------------------------------------

    Minimum Description Length: Developments in Theory and New Applications

            Peter Grunwald, In-Jae Myung, Mark Pitt
            http://quantrm2.psy.ohio-state.edu/injae/workshop.htm

        Inductive inference, the process of inferring a general law from
        observed instances, is at the core of science. The Minimum
        Description Length (MDL) Principle, which was originally proposed
        by Jorma Rissanen in 1978 as a computable approximation of
        Kolmogorov complexity, is a powerful method for inductive
        inference. The MDL principle states that the best explanation
        (i.e., model) given a limited set of observed data is the one that
        permits the greatest compression of the data. That is, the more we
        are able to compress the data, the more we learn about the
        underlying regularities that generated the data. This
        conceptualization originated in algorithmic information theory
        from the notion that the existence of regularities underlying data
        necessarily implies redundancy in the information from successive
        observations. Since 1978, significant strides have been made in
        both the mathematics and application of MDL. For example, MDL is
        now being applied in machine learning, statistical inference,
        model selection, and psychological modeling. The purpose of this
        workshop is to bring together researchers, both theorists and
        practitioners, to discuss the latest developments and share new
        ideas. In doing so, our intent is to introduce to the broader
        NIPS community the current state of the art in the field.

    - -------------------------------------------------------------------------

    Multi-sensory Perception & Learning

            J. Fisher, L. Shams, V. de Sa, M. Slaney, T. Darrell
            http://www.ai.mit.edu/people/fisher/nips01/perceptwshop/description/

        All perception is multi-sensory perception. Situations where animals
        are exposed to information from a single modality exist only in
        experimental settings in the laboratory. For a variety of reasons,
        research on perception has focused on processing within one sensory
        modality. Consequently, the state of knowledge about multi-sensory
        fusion in mammals is largely at the level of phenomenology, and the
        underlying mechanisms and principles are poorly understood. Recently,
        however, there has been a surge of interest in this topic, and this
        field is emerging as one of fast growing areas of research in
        perception.

        Simultaneously and with the advent of low-cost, low-power
        multi-media sensors there has been renewed interest in automated
        multi-modal data processing. Whether it be in an intelligent room
        environment, heterogenous sensor array or the autonomous robot, robust
        integrated processing of multiple modalities has the potential to
        solve perception problems more efficiently by leveraging complementary
        sensor information.

        The goals of this workshop are to further the understanding of the
        both the cognitive mechanisms by which humans (and other animals)
        integrate multi-modal data as well as the means by which automated
        systems may similarly function. It is not our contention that one
        should follow the other. It is our contention, that researchers in
        these different communities stand to gain much through interaction
        with each other. This workshop aims to bring these researchers
        together to compare methods and performance and to develop a common
        understanding of the underlying principles which might be used to
        analyze both human and machine perception of multi-modal
        data. Discussions and presentations will span theory, application, as
        well as relevant aspects of animal/machine perception.

        The workshop will emphasize a moderated discussion format with
        short presentations prefacing each of the discussions. Please
        see the web page for some of the specific questions to be addressed.

    - -------------------------------------------------------------------------

    Neuroimaging: Tools, Methods & Modeling

            B. M. Bly, L. K. Hansen, S. J. Hanson, S. Makeig, S. Strother
            http://psychology.rutgers.edu/Users/ben/nips2001/nips2001workshop.html

        Advances in the mathematical description of neuroimaging data are
        currently a topic of great interest. Last June, at the 7th Annual
        Meeting of the Organization for Human Brain Mapping in Brighton
        UK, the number of statistical modeling abstracts virtually
        exploded (30 abstracts were submitted on ICA alone.) Because of
        its high relevance for researchers in statistical modeling it has
        been the topic of several NIPS workshops. Neuroinformatics is an
        emerging research field, which besides a rich modeling activity
        also is concerned with database and datamining issues as well as
        ongoing discussions of data and model sharing. Several groups now
        distribute statistical modeling tools and advanced exploratory
        approaches are finding increasing use in neuroimaging labs. NIPS
        is a rich arena for multivariate and neural modeling, the
        intersection of Neuroimaging and neural models is important for
        both fields.

        This workshop will discuss the underlying methods and software
        tools related to a variety of strategies for modeling and
        inference in neuroimaging data analysis (Morning, Day 1.)
        Discussants will also present methods for comparison, evaluation,
        and meta-analysis in neuroimaging (Afternoon, Day 1.) On the
        second day of the workshop, we will continue the discussion with a
        focus on multivariate strategies (Morning, Day 2.) The workshop
        will include a discussion of hemodynamic and neural models and
        their role in mathematical modeling of neuroimaging data
        (Afternoon, Day 2). Each session of the two-day workshop will
        include discussion. Talks are intended to last roughly 20 minutes
        each, followed by 10 minutes of discussion. At the end of each
        day, there will be a discussion of themes by all participants,
        with the presenters acting as a panel.

    - -------------------------------------------------------------------------

    Foundations of Occam's razor and parsimony in learning

            David G. Stork
            http://www.rii.ricoh.com/~stork/OccamWorkshop.html

        "Entia non sunt multiplicanda praeter necessitatem"
                      -- William of Occam (1285?-1349?)

        Occam's razor is generally interpreted as counselling the use of
        "simpler" models rather than complex ones, fewer parameters rather
        than more, and "smoother" generalizers rather than those that are
        less smooth. The mathematical descendents of this philosophical
        principle of parsimony appear in minimum-description-length,
        Akaike, Kolmogorov complexity and related principles, having
        numerous manifestations in learning, for instance regularization,
        pruning, and overfitting avoidance. For a given quality of fit to
        the training data, in the absence of other information should we
        favor "simpler" models, and if so, why? How do we measure
        simplicity, and which representation should we use when doing so?
        What assumptions are made -- explicitly or implicitly -- by these
        methods and when are such assumptions valid? What are the minimum
        assumptions or conditions -- for instance that by increasing the
        amount of training data we will improve a classifier's performance
        -- that yield Occam's razor? Support Vector Machines and some
        neural networks contain a very large number of free parameters,
        more than might be permitted by the size of the training data and
        in seeming contradiction to Occam's razor; nevertheless, such
        classifiers can work exceedingly well. Why? Bayesian techniques
        such as ML-II reduce a classifier's complexity in a data-dependent
        way. Does this comport with Occam's razor? Can we characterize
        problems for which Occam's razor should or should not apply? Even
        if we abandon the search for the "true" model that generated the
        training data, can Occam's razor improve our chances of finding a
        "useful" model?

        It has been said that Occam's razor is either profound and true,
        or vacuous and false -- it just isn't clear which. Rather than
        address specific implementation techniques or applications, the
        goal of this workshop is to shed light on, and if possible
        resolve, the theoretical questions associated with Occam's razor,
        some of the deepest in the intellectual foundations of machine
        learning and pattern recognition.

    - -------------------------------------------------------------------------

    Quantum Neural Computing

            Elizabeth Behrman

    - -------------------------------------------------------------------------

    Variable and Feature Selection

            Isabelle Guyon, David Lewis
            http://www.clopinet.com/isabelle/Projects/NIPS2001/

        Variable selection has recently received a lot of attention from
        the machine learning and neural network community because of its
        applications in genomics and text processing. Variable selection
        refers to the problem of selecting input variables that are most
        predictive of a given outcome. Variable selection problems are
        found in all machine learning tasks, supervised or unsupervised
        (clustering), classification, regression, time series prediction,
        two-class or multi-class, posing various levels of challenges. The
        objective of variable selection is two-fold: improving the
        prediction performance of the predictors and providing a better
        understanding of the underlying process that generated the
        data. This last problem is particularly important in biology when
        the process may be a living organism and the variables gene
        expression coefficient. One of the goals of the workshop is to
        explore alternate statements of the problem, including: (i)
        discovering all the variables relevant to the concept (e.g. to
        identify all candidate drug targets) (ii) finding a minimum subset
        of variables that are useful to the predictor (e.g. to identify
        the best biomarkers for diagnosis or prognosis). The workshop will
        also be a forum to compare the best existing algorithms and to
        discuss the organization of a potential competition on variable
        selection for a future workshop. Prospective participants are
        invited to submit one or two pages of summary. Theory, algorithm,
        and application contributions are welcome. After the workshop, the
        participants will be offered the possibility of submitting a full
        paper to a special issue of the Journal of Machine Learning
        Research on variable selection. Deadline for submission: October
        15, 2001. Email submissions to: Isabelle Guyon at
        isabelle@clopinet.com.

    - -------------------------------------------------------------------------

    New Methods for Preference Elicitation

            Craig Boutilier, Holger Hoos, David Poole (chair), Qiang Yang
            http://www.cs.ubc.ca/spider/poole/NIPS/Preferences2001.html

        As intelligent agents become more and more adept at making (or
        recommending) decisions for users in various domains, the need for
        effective methods for the representation, elicitation, and
        discovery of preference and utility functions becomes more
        pressing. Deciding on the best course of action for a user
        depends critically on that user's preferences. While there has
        been much work on representing and learning models of the world
        (e.g., system dynamics), there has been comparatively little
        similar research with respect to preferences. The need to reason
        about preferences arises in electronic commerce, collaborative
        filtering, user interface design, task-oriented mobile robotics,
        reinforcement learning, and many others. Many areas of research
        bring interesting tools to the table that can be used to tackle
        these issues: machine learning (classification, reinforcement
        learning), decision theory and control theory (Markov decision
        processes, filtering techniques), Bayesian networks and
        probabilistic inferences, economics and game theory, among
        others. The aim of this workshop is to bring together a diverse
        group of researchers to discuss the both the practical and
        theoretical problems associated with effective preference
        elicitation and to highlight avenues for future research.

        The deadline for extended abstracts and statements of interest is
        October 19.



    This archive was generated by hypermail 2b29 : Thu Oct 11 2001 - 08:34:54 PDT