Call for Papers
(expired)
Workshop on
Development of Representations

Description

The representation of a learning problem has long been known to be a major factor in learning performance. The nature of appropriate representations and representational change as a part of the learning process have been studied in a variety of forms in a number of subfields within machine learning, artificial intelligence and, more recently, other communities. Despite this fact, representations are typically hand-coded rather than acquired automatically. The goal of this workshop is to explore problems in the area of automated development of representations and to build ties between the various relevant communities.

From a pragmatic standpoint, many learning problems of interest involve extremely large search spaces when viewed in terms of their basic input features. Examples include learning useful behavior for a robot that receives a continuous stream of video input, or learning to play the game of Go. For such problems, an unbiased search is infeasible, and a bias must be employed that focuses the search within the input space so that the size of the problem is effectively reduced. Letting representations develop as part of learning may be viewed as a way of establishing such a bias.

Submissions are encouraged on issues including, but not limited to:

• How can large search spaces be reduced by introducing a helpful bias? Existing approaches relevant to this include bias learning and learning to learn.

• How can related problems become a source for helpful biases? This question is studied in multi-task learning, sequential learning, many-layered learning, and lifelong learning.

• Architectures for variable representations.

• How may representations be used, and when searching the space of representations, what should their evaluation function be? At the revival of neural network research in the mid 1980's, it became clear that internal representations can be learned based on a global feedback signal. However, while this signal is appropriate as an evaluation for a complete system, the representations such systems employ may require a different evaluation:

  • assessing modularity: Can a representation be used in multiple contexts? Structural vs. functional modularity.

  • assessing value: How useful is the information a representation extracts to the construction of solutions? This is a credit assignment question, and recent work on establishing stable economies of value may shed new light on this.

• Statistical techniques for assessing modularity. The modularity of a representation relates to a reduced dependency on elements that are not part of the representation.

• Bayesian techniques for learning representations.

• The relationship between statistical techniques and other approaches to credit assignment.

• Hierarchy. The size of input spaces than can be handled may be scaled up by constructing representations from existing representations, leading to a hierarchy of representations.

• Practical methods for hierarchical Bayesian inference.

• Extracting symbols from sensors. How can raw sensor information be used to extract compact representations or symbols?

• How may representations and the solutions employing them be developed simultaneously? One approach to this question is studied in the sub-discipline of evolutionary computation known as co-evolution.

• Methods for constructive induction.

• Development of theoretical terms through, for example, predicate invention.

• Emerging issues in evolutionary and computational biology on the importance of change of representation in gene expression.

• Change of representation that occurs over the lifetime of an embedded agent.

Format

The workshop will be organized so as to maximize interaction, discussion, and exchange of ideas. The day will start with an invited talk and will be followed by a series of paper presentations grouped by topic. Each presentation will be short, e.g. 10 or 15 minutes, with 5 minutes allotted to questions on the content of the talk. At the end of each group of papers the presenters will participate in a panel discussion to answer questions of a more general sort related to the topic and the relationship between the papers in that group. We will include a panel discussion on emerging problems in the area of development of representation, and conclude the day by inviting all participants to join in an open discussion with the goal of identifying the main themes of the day and establishing a research agenda.

Submissions

Submissions may be either a full technical paper (up to 8 pages) or a position statement in the form of an extended abstract (one or two pages). Electronic submissions (PostScript, PDF, or HTML) are preferred and should be sent by April 22 to either of the co-chairs (Edwin de Jong at edwin@cs.brandeis.edu or Tim Oates at oates@cs.umbc.edu). Please format your submission according to the ICML-2002 formatting guidelines.

Important Dates

Organization

Edwin de Jong
Computer Science Department
Brandeis University MS018
Waltham, MA 02454-9110
1.781.736.3366
edwin@cs.brandeis.edu

Tim Oates
CSEE Department
University of Maryland Baltimore County
1000 Hilltop Circle
Baltimore, MD 21250
1.410.455.3082
oates@cs.umbc.edu

Program committee

Jonathan Baxter	WhizBang! Labs
Rich Caruana	Cornell
Rod Grupen	University of Massachusetts, Amherst
Tom Heskes	University of Nijmegen, The Netherlands
Leslie Kaelbling	MIT
Justus Piater	INRIA Rhone-Alpes, France
Jude Shavlik	University of Wisconsin, Madison
Paul Utgoff	University of Massachusetts, Amherst