Connectionists: CFP: NIPS 2005 Workshop: Nonparametric Bayesian methods

[Matthew Beal]

################################################################

                CALL FOR PARTICIPATION

           Open Problems and Challenges for
            Nonparametric Bayesian methods
                 in Machine Learning

                  a workshop at the

       2005 Neural Information Processing Systems
                 (NIPS) Conference

     Friday, December 9, 2005, Whistler BC, Canada

  http://aluminum.cse.buffalo.edu:8080/npbayes/nipsws05

   Deadline for Submissions: Monday, October 31, 2005
   Notification of Decision: Friday, November 4, 2005

################################################################

Organizers:
-----------
Matthew J. Beal, University at Buffalo, State University of New York
Yee Whye Teh, National University of Singapore

Overview:
---------
Two years ago the NIPS workshops hosted its first forum for discussion
of Nonparametric Bayesian (NPB) methods and Infinite Models as used in
machine learning. It brought to bear techniques from the statistical
disciplines on perennial problems in the NIPS community, such as model
size and structure selection from a Bayesian viewpoint.

NPB methods include some topics that are heavily studied in the NIPS
community such as Gaussian Processes and more recently Dirichlet
Process Mixture models, and other topics that are only now starting to
make an impact. NPB methods are attractive to machine learning
practitioners because they are both powerful and flexible: the key
ingredient in an NPB formalism is to side-step the traditional
practice of parameter fitting by instead integrating out the possibly
complex parameters of the model, which allows interesting situations
to exist in the model such as (countably) infinite components in a
mixture model, infinite topics in a topic model, or infinite
dimensions in a hidden state space; also, the flexibility of an NPB
model is often very useful in domains where it is difficult to
articulate priors or likelihood functions, such as in text and
language modeling, spatially-dependent process modeling, etc.

Since the last workshop, models such as Sparse Factor Analysis, Latent
Dirichlet Allocation, Hidden Markov Models, and even robot mapping
tasks have all benefitted from the flexibility of a nonparametric
Bayesian approach, and these nonparametric alternatives have been
shown to give superior generalization performance, as compared to
finite model selection techniques.

It is time now after two years to collect together these various
research directions, and use them to define and delineate the
challenges facing the nonparametric Bayesian community, and with this
the set of open problems that stand a reasonable chance of being
solved with focussed research plans. The workshop will focus less on
well-studied topics like Gaussian Processes and more on potential new
ideas from the statistics community. To this end, we will have a
number of experts on nonparametric Bayesian methods from statistics to
share their experiences and expertise with the general NIPS community,
in an effort to transfer and build upon key methodologies developed
there, since the time is ripe for the two groups to coalesce.

In particular, we would like the workshop to address:

* New techniques: What techniques and methodologies are currently
being used in the statistics communities, and which of these can be
transferred to be used in machine learning applications? Conversely,
are the techniques developed withing the NIPS community of interest to
the general statistics communities?

* New problems: There are still a wide variety of problems in the NIPS
community that cannot be elegantly solved by nonparametric Bayesian
models, for example, problems needing smoothly time-varying or
spatially-varying processes. It would be useful to identify these
problems and the necessary characteristics of any nonparametric
solutions. This can serve as concrete goals for further research.

* Computational/Inferential issues: Inference in nonparametric models
is for the most part carried out using expensive MCMC sampling, but
recently variational and Expectation Propagation methods have been
applied to isolated cases only. For more popular use of these models
we need more efficient and reliable inference schemes. Can these
methods scale to high dimensional data, and to large databases such as
email repositories, news reports, and the world-wide-web?  Could it be
that NPB methods are just too much work for too little benefit?

Format:
-------
This is a one-day workshop, designed to be highly interactive,
consisting of 3 or 4 themed sessions with short talks and lengthy
moderated discussion periods.  We anticipate a strong response and
will likely have a poster session in between the morning and afternoon
sessions. We have attracted several statisticians from outside the
NIPS beaten track and as such we will also have a Distinguished Panel
of statisticians/machine learners to discuss the points arising during
the workshop's discussions.

Call for Contributions:
-----------------------
Potential speakers/discussants/attendees are encouraged to submit
(extended) abstracts of 2-4 pages in length outlining their research
as it relates to the themes above, before *Monday, October 31*.  We
are looking for position papers, extensions of theory and
applications, as well as case studies of nonparametric methods.  If
there is overwhelming response we will accommodate a poster session in
the afternoon break in between morning and evening sessions.  All
chosen abstracts will be posted on the workshop website beforehand to
stimulate discussion.

  Deadline for Submissions: Monday, October 31, 2005
  Notification of Decision: Friday, November 4, 2005
  Date of workshop: Friday, December 9, 2005

Please email your submissions to

  mbeal at cse.buffalo.edu  and/or  yeewhye at gmail.com

We encourage you to log in, post questions, and contribute to the
pages of the workshop website, at

  http://aluminum.cse.buffalo.edu:8080/npbayes/nipsws05

which is in Wiki format, either in person or anonymously.  Once you
have a login if you wish to modify the pages please contact us for
permissions.  The website is intended to serve as a resource for you
as nonparametric Bayesians, both before and after the workshop, and
already contains plenty of links to literature on NPB from within and
outside of the NIPS community.  With your input we can tailor the
workshop according to your suggestions.  Also, for you to give
comprehensive feedback on the topics to be covered, we have provided a
survey form at

  http://aluminum.cse.buffalo.edu:8080/npbayes/nipsws05/survey

which you are free to fill out only partially, and anonymously if
desired.

Thank you

-Matt Beal & Yee Whye Teh