[Imageworld] NIPS05 Interclass Transfer Workshop: why learning to recognize many objects might be easier than learning to recognize just one?

From: Michael Fink (fink@cs.huji.ac.il)
Date: Sun Sep 25 2005 - 10:59:06 CEST

Call for Papers

NIPS*05 Workshop on interclass transfer:
Why learning to recognize many object classes might
be easier than learning to recognize just one


NIPS 2005
Submission deadline: 21 October
Accept/Reject notification: 05 November


Andras Ferencz, University of California at Berkeley
Michael Fink, The Hebrew University of Jerusalem
Shimon Ullman, Weizmann Institute of Science

Workshop Description

The human perceptual system has the remarkable capacity to recognize
numerous object classes, often learning to reliably classify a novel
category from just a short exposure to a single example. These skills are
beyond the reach of current multi-class recognition systems. The workshop
will focus on the proposal that a key factor for achieving such capabilities
is the use of interclass transfer during learning. According to this view, a
recognition system may benefit from interclass transfer if the multiple
target classification tasks share common underlying structures that can be
utilized to facilitate training or detection. Several challenges follow from
this observation. First, can a theoretical foundation of interclass transfer
be formulated? Second, what are promising algorithmic approaches for
utilizing interclass transfer. Finally, can the computational approaches for
multiple object recognition contribute insights to the research of human
recognition processes?

In the coming workshop we propose to address the following topics:

* Explore the human capabilities for multi-class object recognition and
how these capacities motivate our algorithmic approaches.

* Attempt to formalize the interclass transfer framework and define what can
be generalized between classes (for example, learning by analogy from the
"closest" known category vs. finding useful subspaces from all categories).

* Analyze state-of-the-art solutions aimed at recognizing many objects or at
learning to recognize novel objects form very few examples (e.g. contrasting
parametric vs. non-parametric approaches).

* Characterize the problems in which we expect to observe high transfer
between classes.

* Delineate future challenges and suggest benchmarks for assessing progress

The workshop is aimed at bringing together experimental and theoretical
researchers interested in multi-class object recognition in humans and

Imageworld mailing list

Home   Login   Meetings   Courses   Belgium   Maillist   Credits
  Servers   Links   Archive

Please send comments and suggestions to Mike Wijnants.

Page last updated on © BBF 1998 all rights reserved