Call for papers: NIPS 2009 workshop on The Generative and Discriminative Learning Interface

CALL FOR PAPERS – The Generative and Discriminative Learning Interface
(supported by PASCAL 2)

Workshop at the 23rd Annual Conference on Neural Information
Processing Systems (NIPS 2009)
December 12, 2009, Whistler, Canada

Submission Deadline: October 25, 2009


Generative and discriminative learning are two of the major paradigms
for solving prediction problems in machine learning, each offering
important distinct advantages. They have often been studied in
different sub-communities, but over the past decade, there has been
increasing interest in trying to understand and leverage the
advantages of both approaches. The goal of this workshop is to map out
our current understanding of the empirical and theoretical advantages
of each approach as well as their combination, and to identify open
research directions.


In generative approaches for prediction tasks, one models a joint
distribution on inputs and outputs and parameters are typically
estimated using a likelihood-based criterion. In discriminative
approaches, one directly models the mapping from inputs to outputs
(either as a conditional distribution or simply as a prediction
function); parameters are estimated by optimizing various objectives
related to a loss function. Discriminative approaches have shown
better performance with enough data, as they are better tuned to the
prediction task and are more robust to model misspecification. Despite
the strong empirical success of discriminative methods in a wide range
of applications, when the structures to be learned become complex
(e.g. in machine translation, scene understanding, biological process
discovery), even large training sets become sparse relative to the
task, and this sparsity can only be mitigated if some other source of
information comes into play to constrain the space of fitted models,
such as unlabeled examples, related data sources or human prior
knowledge about the problem. Generative modeling is a principle way of
encoding this additional information, e.g. through probabilistic
graphical models or stochastic grammar rules. Moreover, they provide a
natural way to make use of unlabeled data and can be more
computationally efficient for some models.

See for a more detailed
background with references.

The aim of this workshop is to provide a platform for both theoretical
and applied researchers from different communities to discuss the
status of our understanding on the interplay between generative and
discriminative learning, as well as to identify forward-looking open
problems of interest to the NIPS community. Examples of topics of
interest to the workshop are as follows:
* Theoretical analysis of generative vs. discriminative learning
* Techniques for combining generative / discriminative approaches
* Successful applications of hybrids
* Empirical comparison of generative vs. discriminative learning
* Inclusion of prior knowledge in discriminative methods
(semi-supervised approaches, generalized expectation criteria,
posterior regularization, etc.)
* Insights into the role of generative / discriminative interface for
deep learning
* Computational issues in discriminatively trained generative
models/hybrid models
* Map of possible generative / discriminative approaches and combinations
* Bayesian approaches optimized for predictive performance
* Comparison of model-free and model-based approaches in statistics or
reinforcement learning


Dan Klein, UC Berkeley

Tony Jebara, Columbia University

Phil Long, Google

Ben Taskar, University of Pennsylvania

John Winn, Microsoft Research Cambridge


Deadline for abstract submission: October 25, 2009
Notification of acceptance: November 5, 2009
(NIPS early registration deadline is November 6)
Final version: November 20, 2009
Workshop: December 12, 2009


Westin Resort and Spa / Hilton Whistler Resort and Spa
Whistler, B.C., Canada


Researchers interested in presenting their work and ideas on the above
themes are invited to submit an extended abstract of 2-4 pages in pdf
format using the NIPS style available at (author names don’t need to
be anonymized). Submissions will be accepted either as contributed
talks or poster presentations, and we expect the speakers to provide a
final version of their paper by November 20 to be posted on the
workshop website.

Sign on at:
to submit your paper (you’ll need to create a login first).


This 1 day workshop will have a mix of invited talks (3), contributed
talks (4-8), a poster session as well as a panel discussion. We will
leave plenty of time and encourage discussion throughout the day.

We also encourage the participants to visit the online forum in
December to discuss the submitted papers and the themes of the

SPONSOR: PASCAL 2 (non-core workshop)..


Simon Lacoste-Julien (University of Cambridge)
Percy Liang (UC Berkeley)
Guillaume Bouchard (Xerox Research Centre Europe)


gen.disc.nips09 at