[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
NIPS*99 Workshop Announcements
From: Sue Becker <becker@wvfUZLTFKgPFOhYv_bD5L3qLEs8XlF6ZPeagwClnAEYdNhsLvZu_0Yya5W_t_0aenrhZi-s6CaIyHJqi5NS7cJSbqBvL3dDsS0w.yahoo.invalid>
Dear Connectionists,
Below are announcements for the 16 workshops to be held at the post-NIPS
meeting in Breckenridge, December 2-4. Please refer to the NIPS workshops
web page and links to individual workshops for further details:
http://www.cs.cmu.edu/Groups/NIPS/NIPS99/Workshops
The scheduling of workshops to particular days will be completed soon and
will be added to the above web page when it's ready.
Sue Becker and Rich Caruana, NIPS*99 workshop co-chairs
---------------------------------------------------------------------
COMPLEXITY AND NEURAL COMPUTATION:
THE AVERAGE AND THE WORST CASE
http://www.dsi.unive.it/~nips99
Organizers:
Marco Budinich (Trieste, Italy)
Marcello Pelillo (Venice, Italy)
NP-complete problems have attracted many studies, also in the neural
network community. Recent advances obtained both from the analysis of the
average case complexity (in contrast to the, standard, worst case
analysis) and from the design of dynamical systems that tackle NP-complete
problems could benefit from a substantial cross-fertilization. For
example, the analysis of the average case gives possible indications for
the design of new algorithms that could be esaily implemented with
neural-networks like techinques. This (informal) workshop is devoted
to the exchange of ideas and experiences between people tackling
NP-complete problems (both with standard and non-standard techniques,
e.g. neural networks) and researchers in the field of the average
complexity.
---------------------------------------------------------------------
Neural Mechanisms of Music Processing
URL: http://bach.ece.jhu.edu/nips99/
Organizers: Tim Edwards, Shihab Shamma, Malcolm Slaney
The term "music" covers a large variety of signals, from the simple
to the complex. It is easy to choose stimuli to study a particular
phenomenon, be it pitch detection, harmonic analysis, sequence
prediction, or spatial location. However, mixtures of time-varying
harmonics, intertwining sequences, resonances and spatial reflections
easily make music as complex and difficult to analyze as speech.
What makes music especially interesting is the ability to control the
complexity of data, allowing independent control over a variety of
parameters which are not possible to separate in speech. This
workshop will concentrate on several key issues which are currently
prominent in acoustics research. In particular, the intent of the
workshop is to bring together areas of research which appear to be
strongly related to the problem of music processing and to each other.
---------------------------------------------------------------------
Learning Relational Data Representations : Call for papers
http://www.cs.york.ac.uk/~wilson/workshop/main.html
Organizers: R. Wilson (University of York, England), B. Huet (Eurecom, France)
Graph structures play a critical role in many areas of pattern recognition
and machine intelligence, particularly in the domains of reasoning and
recognition. These structures offer considerable advantages in terms of
flexibility and representational power, in contrast to the Euclidean feature
vectors so often adopted in the field of neural computation. However, the
powerful analytical tools, which are available for traditional feature
vectors, are more difficult to apply to relational structures. The study of
learning theory applied to relational graph structures will provide
important insights for both the neural networks and graphical modeling
communities. It is hoped that such methods will provide a bridge between the
two bodies of work. Appropriately modified learning techniques such as
statistical learning, Bayesian learning, graph clustering and classification
will increasingly be of importance. Of particular interest is the task of
learning optimum relational representations of data.
---------------------------------------------------------------------
Call for Participation and Submission:
USING UNLABELED DATA FOR SUPERVISED LEARNING
http://stat.cs.tu-berlin.de/nips99/
Thore Graepel
Ralf Herbrich
Klaus Obermayer
Department of Computer Science
Technical University of Berlin
In many applications of supervised learning an abundance of unlabeled
data is available while labeled instances are scarse. This situation
arises, e.g., in speech and character recognition, in data-mining
tasks, or in the classification of World Wide Web pages. The
availability of unlabeled data clearly poses the challenge of how to
use them in order to improve generalization in supervised learning. To
this end different approaches have been put forward: Unlabeled data
can be used to control the complexity of function classes, they form
the basis for query filtering, and even Transduction can be
viewed as a way of using unlabeled data in the working set. Also, some
people argue that nothing at all can be learnt from unlabeled data in
a supervised setting. Finally, using unlabeled data for supervised
learning constitutes a convergence of supervised and unsupervised
learning and should benefit from the confluence of knowledge from both
these fields. The proposed workshop can be seen as a continuation of
last year's NIPS-workshop "Integrating Supervised and Unsupervised
Learning", however with considerably different focus. Note, that we
encourage submissions for 10-15 min. presentations of relevant
work (see web page above).
---------------------------------------------------------------------
CALL FOR PARTICIPATION ---- NIPS*99 Post Conference Workshop:
"Neural mechanisms of perceptual selection in visual and prefrontal cortex"
Date & Venue: December 3, 1999. Breckenridge, Colorado.
Organizers: Erik D. Lumer
Wellcome Dept. Cognitive Neurology, Institute of Neurology,
University College London, 12 Queen Square, London WC1N 3BG
Geraint Rees
Division of Biology 139-74,California Institute of Technology,
Pasadena, CA 91125
Workshop URL: http://www.fil.ion.ucl.ac.uk/nips
Summary:
This workshop will bring together experimental and theoretical leaders in the
fields of perceptual organization, attention, and perceptual decision, with
the explicit goal of laying out a framework for understanding how the modulatory
effects of perceptual selection observed in visual cortex are produced. In
particular, how are the influences of higher visual and nonvisual areas relayed
to earlier stages of visual processing? The workshop will comprise two parts.
First, a relatively small number of invited speakers will give presentations of
recent work, each followed by discussion. In addition, there will be a highly
interactive panel and general discussion of alternative accounts of and future
plans to elucidate the neural mechanisms of perceptual selection.
Confirmed speakers include: Jochen Braun, Maurizio Corbetta, Sabine Kastner,
Victor Lamme, Rajesh Rao, Mike Shadlen.
---------------------------------------------------------------------
ADVANCED MEAN FIELD METHODS
http://www.ncrg.aston.ac.uk/nips99/workshop.html
Organizers: Manfred Opper and David Saad
(NCRG, Aston University, Birmingham, U.K.)
Mean field methods provide tractable approximations in high dimensional
probabilistic models for which exact computations are not feasible.
They have found widespread applications especially in the growing field
of graphical models. The workshop focusses on the theoretical foundations
and the applications of advanced mean field approaches which go beyond simple
approximations by factorizable distributions. Topics will include e.g. recent
developments in the variational method, belief propagation techniques and
approaches that are motivated by ideas from Statistical Physics
(e.g. 'TAP' mean field theory).
---------------------------------------------------------------------
Call for participation :
"NEURAL NETWORKS AND HUMAN MOVEMENTS SIMULATION: NEW FRONTIERS"
http://neuropent.polito.it/events/events01.html
Organizer: Prof. Eros Pasero
Dip. Elettronica - Politecnico di Torino
c.so Duca d. Abruzzi 24 - 10129 Torino - Italy
Ph: +39 011 5644043 Fax: +39 011 5644099
Email: pasero@SBlBH-mJRvPH1xYk4wUPUhZC3udj8He596HZRnZFMufHl3W6PKvJ9rJyeBD_CfvhK7h5Tdhi1rE.yahoo.invalid
Summary:
Ergonomics is becoming a major topic in designing new objects, from a
can-opener to the cockpit of a new car, from a keyboard to an assembly line
for a car manufacturer. Whereas it's quite simple to build a can-opener
sample, it could be very expensive to build the cockpit of a new car. The
virtual reality can give a significant help to this phase. By means of
powerful graphical computers you can design your object and simulate its
ergonomic behavior. But how can you simulate the interactions between humans
and virtual objects? Software simulators using powerful algorithms to
replicate human movements can be used to insert an artificial mannequin in a
3D-CAD environment. Unfortunately movements generated by algorithms are
sometimes more similar to robotic movements than to human movements.
Nevertheless the "fuzzyness" of human choices and reasoning are difficult to
simulate with models and algorithms. If you ask to ten people to catch a pen
on a table usually you can observe ten different movements. This "free will"
is not easy to simulate! Artificial Neural Networks can be a promising answer
to these problems. Why not to use artificial brains to simulate real brains?
Today results will be presented and discussed in this workshop. Suggested
arguments are:
a) The organization of CAD tools for 3D design and modelization
b) The problem of the animation of artificial mannequins
c) How to transport human biomodels to virtual models.
d) How to train neural networks with actual human movements and how to use
it inside virtual reality.
e) Industrial applications
---------------------------------------------------------------------
Call for Participation: Geometric Methods in Learning
(Shun-ichi Amari (RIKEN-Japan), Amir Assadi, UW-Madison (Chair), Tommy
Poggio (MIT), Pawan Sinha (MIT))
http://www.cms.wisc.edu/~cvg/nipsevents.html
Throughout the history of science, geometric methods have played a key role
in the development of scientific theory, in at least two very influential
ways:
a) as a vehicle for communication of profound ideas about nature;
b) through use of the logical precision of geometric methods and powerful
tools to further explore new ideas and to establish far reaching theories that
suggest new experiments. Two famous examples of this are seen in gravitation
and relativity. The time is ripe to attract the attention of the learning
community to geometric methodsand to take on an endeavor to:
1.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. Among the discussion topics,
we envision the following: Information geometry, differential topological
methods for turning local estimates into global quantities and invariants,
Riemannian geometry as a framework to explore nonlinearity, spatial
statistics, impact of learning theory on future development of geometry,
and examples of how theoretical physics has developed their own version of
geometric tools to formulate unification theories.
---------------------------------------------------------------------
CALL FOR PARTICIPATION
Spike Timing and Synaptic Plasticity
http://www.pitt.edu/~pwm/LTP_LTD_99
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. 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.
The major goals of the workshop are:
* To review current experimental results on spike-timing-dependent
synaptic plasticity.
* To discuss models and mechanisms for this form of synaptic
plasticity.
* To explore its implications for development and learning in neural
networks.
--------------------------------------------------------------------------
MCMC METHODS FOR MACHINE LEARNING
http://svr-www.eng.cam.ac.uk/~jfgf/nips99.html
ORGANISERS: Christophe Andrieu and Arnaud Doucet (Cambridge), Neil Gordon
and Alan Marrs (Defence Research Agency, UK) and Nando de
Freitas (UC Berkeley)
Markov chain Monte Carlo (MCMC) techniques are a set of powerful and
flexible simulation methods that may be applied to solve integration and
optimisation problems in large dimensional spaces. These two types of
problem play a fundamental role in machine learning, Bayesian statistics
and decision analysis. MCMC methods were introduced in the physics
literature in the 1950's and are now at the heart of a recent revolution in
applied statistics. Many of the new advances in MCMC simulation, including
model selection and model mixing, parallel chains, perfect sampling,
forward-backward sampling and on-line MCMC among others, have been
overlooked to a large extent by the neural networks and machine learning
communities. This workshop will attempt to provide a simple tutorial
review of these state-of-the-art simulation-based computational methods.
It will also focus on application domains and encourage audience
participation.
---------------------------------------------------------------------------
LEARNING WITH SUPPORT VECTOR MACHINES: THEORY AND APPLICATIONS
http://lara.enm.bris.ac.uk/cig/nips99.htm
Deadline for receipt of abstracts is **** 20th October 1999 ****.
Organisers; C. Campbell (Bristol, UK), C. Burges (Lucent Technologies,
USA), N. Cristianini (Bristol, UK)
Support Vector Machines have become established as a powerful technique
for solving a variety of classification, regression and density estimation
tasks. This workshop will cover Support Vector Machines and related
techniques based on the concepts of large margins or the use of kernel
methods. Topics covered will include:
- new theoretical contributions
- algorithms and implementation techniques
- applications
Contributors to the workshop will be invited to submit their papers to a
special issue of the journal Machine Learning edited by the workshop
organisers.
---------------------------------------------------------------------
Statistical Learning in High Dimensions
---------------------------------------
Marina Meila and Andrew Moore
URL: http://www.ai.mit.edu/~mmp/workshop-nips99/workshop-nips99.html
As more and more knowledge domains enter the scope of machine
learning, researchers are increasingly often confronted with high and
very high dimensional data. Of the wide spectrum of existing machine
learning methods only a few ones are currently applied to high
dimensional problems. The goal of this workshop is to better
understand the sources of difficulty in training (and using)
high-dimensional statistical models and to expose the participants to
recent solutions and to approaches outside the traditional scope of
NIPS (e.g. multiscale models). Therefore we will bring together
researchers from domains dealing with constructing statistical models
of high-dimensional data (text and image processing, computational
biology, diagnosis systems), from data mining, graphical models and
algorithms to discuss issues that are relevant across different
fields. We will focus on the algorithmic aspects by emphasizing on:
* fast and very fast exact algorithms
* models/training algorithms with no local minima (e.g. support
vectors, tree distributions)
* How to efficiently prune irrelevant features. Implicit versus
explicit feature selection techniques.
* How to use prior knowledge in speeding up training and search?
Lessons from domain specific paradigms.
* Supervised versus unsupervised training. Often times, one finds
that density estimators perform well in classification/recognition
tasks. What causes this behaviour? Are there lessons to be learned
that would improve classifier training?
Contact Info
Marina Meila
Smith Hall 208
Carnegie Mellon University
5000 Forbes Avenue
Pittsburgh PA 15213
mmp@XvOZcmfV2yEkGzkR7Wqt8HlEINwBoHek8z3ugmuL56ILaGN6pCY4JI2YPbF1YG2HwxFoZBid3m_S.yahoo.invalid
Phone:(412)268-8424
Andrew W. Moore
Smith Hall 221
Carnegie Mellon University
5000 Forbes Avenue
Pittsburgh PA 15213
awm@ydHjyT2gN-hAxlcAWpQaBZiBu48t-wr9W10fC3MPCBaO5ssP8mT-Y3MPelfWUdF2jxDhwRk.yahoo.invalid
Phone:(412)268-7599
Fax:(412)268-5571
---------------------------------------------------------------------
CONTEXT AND ADAPTATION IN STATISTICAL THEORIES OF VISION
http://vision.psych.umn.edu/www/people/schrater/NIPS*99workshop.htm
Organizers: P. Schrater (U. Minnesota), E. Simoncelli (NYU, New York),
D. Kersten (U. Minnesota)
Although a great deal of work has recently been done on natural image
statistics,the primary focus has been on understanding the low-order ensemble
statistics of images. These ensemble statistics are assumed to be collected
over long time intervals, and thus over a number of distinct contexts.
Because statistical properties in a scene tend to be more homogeneous over
shorter time intervals, there may be advantages to systems which do short time
statistical analyses, especially for short duration events such as motion
processing and tasks requiring online control such as the visual control of
reaching. The goal of the workshop is to bring together researchers
interested statistical approaches to vision involving adaptive, short-time, or
online processes. Topics that may be addressed include: theories of context
and context learning, theories of perceptual learning and adaptation,
statistical interpretations of cortical dynamics, adaptive statistical
processing (e.g. sequential parameter estimation, dynamic Bayes Nets, Kalman
filtering, EM, etc.), visuo-motor control, and the perception of dynamics.
___________________________
Paul Schrater, Ph.D.
Psychology Dept, U. of Minn.
N218 Elliott Hall
75 East River Road
Minneapolis, MN
USA 55455
Phone: 612-626-8638 FAX: 612-626-2079
email: schrater@-crwV0V63LMD7HTXUWTbGSPFOtCOAX8-w_dG8j7smlt_fGm1O4UFKAf5wukM-GJNhXw3sdz_anJW9JqUyoFjdQ.yahoo.invalid
web: http://vision.psych.umn.edu/www/people/schrater
---------------------------------------------------------------------
Neuronal Response Variability - Curse or Blessing?
Organizer: Klaus Pawelzik
The subject of this workshop is the high degree of variability of
neuronal responses as measured in the cortex. This robust phenomenon
challenges theories of neuronal coding which are based on
spatio-temporally coordinated activities (Softkey and Koch,
Aertsen) and could provide evidence for a rather sloppy rate code
(Shadlen and Newsome).
Both opposing views have been hotly debated in neuroscience during
the past and now after some dust has settled it is a good time to take
a fresh view on this subject which lies at the heart of our
understanding how brains process information. The purpose of this
workshop is to present the opposing views to a broadly interested
audience and to discuss their evidence and implications. Obviously a
significant source of variability is given by the huge amount of ongoing
activity in the cortex. Optical imaging with voltage sensitive dyes
(A.Arielli et al.) and dual intracellular recordings in vivo (I.Lampl et
al.) have revealed a surprising amount of synchronous subthreshold
activity. While in the past the ongoing activity has been treated
largely as a kind of noise that masks the stimulus dependent part
of neuronal responses this workshop will also discuss a novel hypothesis
which suggest that the subthreshold activity is a signal generated
within the cortex that represents contextual information and can be
effective as a selective gain control (A. Aertsen, M.Bethge).
---------------------------------------------------------------------
Computational Brain Imaging: Beyond Modern Phrenology
http://www.cnl.salk.edu/~giedrius/nips99/index.html
Organizers: Geoff Boynton, Giedrius Baracas, The Salk Institute
The goal of the proposed workshop is to explore how functional brain
imaging research can benefit from a quantitative, rather than
qualitative approach to experimental design and data analysis.
The majority of functional MRI (fMRI) studies employ a localizationist
approach which attempt to locate specific brain centers associated
with various perceptual, motor, and cognitive functions. Typically,
results are shown as thresholded statistical parameter maps (such as
Z-score maps) which show the foci of "brain centers" associated with a
specific stimulus or task. It has been noted, however (Wandell,
1999), that the increase in signal-to-noise ratio with the advent of
more advanced imaging technology can only increase the number of brain
regions that are activated above a given significance threshold, thus
making the notion of a "cognitive brain center" untenable.
In contrast, recent approaches are exploiting the continuous,
quantitative nature of the raw fMRI signal to detect subtle
covariations between the fMRI signal and various stimulus parameters.
Several groups are now employing this quantitative approach by
exploring the mapping between the fMRI signal and parameters such as
position in the visual field (mapping retinotopy), stimulus strength
and duration, and the behavioral performance of psychophysical tasks.
The proposed workshop will address issues related to the relationship
between the underlying neuronal response and the fMRI signal. This
topic is timely; functional MRI is nearing its 10-year anniversary,
and fMRI research can no longer rely on the novelty of the technique.
It is time to take advantage of recent technological and theoretical
advances to move beyond basic localization studies and toward
quantitative studies of brain function.
---------------------------------------------------------------------
Overcomplete representations and
nonlinear approaches to ICA
http://www.cnl.salk.edu/~tewon/ICA/nips_workshop.html
Organizers: Te-Won Lee (The Salk Institute)
Kenneth Kreutz-Delgado (UCSD)
Michael Lewicki (CMU)
Recently, there has been a great interest in statistical models for
learning overcomplete representations. A popular method for learning
data representations is Independent Component Analysis (ICA). Although
most ICA research was focused on complete representations, we believe
that statistical models with overcomplete representations will lead to
new powerful techniques for signal- and image processing. There is
also a link between overcomplete representations and efficient sensory
coding in the brain which shows the vitality of the emerging new
perspectives. An example is Hubel and Wiesel's finding of oriented
receptive fields in cat visual cortex which has been substantiated by
coding natural scenes (Olshausen, Bell, and Lewicki) using
complete/overcomplete representations. The methods derived by
considering statistical and information theoretic principles such can
be used for coding images and audio signals which are more efficient
than other traditional engineering methods.
The goal of this workshop is to bring together researchers who are
interested in exchanging their ideas and exploring new theories for
overcomplete representations. We anticipate talks on advanced topics
in overcomplete representations, mixture models, nonlinear independent
component analysis, information theoretic coding, image understanding,
speech modeling, and pattern recognition using overcomplete
representations.
---------------------------------------------------------------------