Boulder Computational Learning Group
Spring 1998 Schedule

Unless otherwise noted, meeting time is on Mondays, 1-3 p.m., in ECOT 831a (Engineering center, 8th floor, small conference room). Dates in red are at the usual time and place. Dates in blue are special meeting dates.

Schedules from previous semesters

January 12: No Meeting (first day of classes)

January 19: No Meeting (holiday)

January 26: Satinder Singh

Recent research on reinforcement learning

February 2: Tor Mohling

A computational model of unconscious and conscious processes in human anagram solving

February 9: Marc Pickett

February 16: NO MEETING (President's day)

March 2: Manav Misra

Graph matching for visual pattern recognition

March 6: Larry Reeker, NSF (tentative)

March 9: Robert Dodier

Belief nets

March 19, 3:45 p.m., ECCR 265: Yoram Singer, AT&T

``80 and tea: ho May I hell Pooh ?'': Efficient algorithms for multi-class multi-label text categorization

When an At&T customer dials 00, she gets an operator who asks: ``AT&T: how may I help you?''. A possible answer might be: ``Ah yes, I would like to call 908-508-1464 and charge it to my Visa.'' Being able to automatically determine what the caller actually wants (in this case a `dial-for-me' request and a `credit-card-call' request) is a difficult task especially since the might be more than one action to be taken due to the caller's request.

This work focuses on algorithms which learn from examples to perform text and speech categorization tasks. We first show how to extend the standard notion of classification by allowing each instance to be associated with multiple labels. We then discuss our approach for text classification which is based on a new and improved family of boosting algorithms. We describe in detail an implementation, called BoosTexter, of the new boosting algorithms for text classification tasks. We present results comparing BoosTexter and a number of other text classification algorithms. We conclude with a demonstration of the system for the ``AT&T: how may I help you?'' speech categorization task.

Joint work with Rob Schapire (AT&T Labs).

March 19, 5 p.m., Room 158, Department of Psychology, University of Denver (2155 S. Race St.): Marie Banich, U. Illinois

The role of the corpus collosum in modulating attention CANCELLED

March 30: Mike Mozer / Randy O'Reilly

Practice talks for Cognitive Neuroscience Conference:

Temporal dynamics of cognition (Mozer)
Principles for Learning and Processing in the Cortex: Reconciling Interactivity and Generalization using Competition and Hebbian Learning (O'Reilly)

April 2, 12:00-1:00 p.m., Muenzinger D430: Ray Jackendoff, MIT

What's in the Lexicon?

A basic psychological question that has not received quite enough attention in the theoretical literature is: What words and uses of words must be stored in long-term memory, and which can be computed online in short-term memory? This question leads to a distinction between two kinds of regularities that have both been called 'lexical rules': productive rules for which one always knows, given an appropriate input, what the output will be; and semiproductive rules, for which one has to know whether the output is an actual word and what its peculiarities might be. I will argue that although the outputs of semiproductive processes must be stored, those of productive processes may be composed online (and in many cases must be). This of course has an important bearing on the hot issue of 'one or two systems' in morphology.

April 6: Michael Littman

This is the regular BCLG meeting, but will be held in Muenzinger D430

April 17, 10-12, ECOT 831a, Peter Marbach

Reinforcement learning theory

April 20: Manav Misra

Modeling task-dependent performance in visual search with reinforcement learning

April 27: John Williams

Experiments with Memoryless Algorithm to Learn Optimal Stochastic Policies for Partially Observable Markov Decision Processes

Partially Observable Markov Decision Problems (POMDPs) are an important class of learning problems which present unique theoretical and computational difficulties. In the absence of the Markov property, classical reinforcement learning techniques may no longer be effective, and memory-based approaches to state-estimation are notoriously computationally intensive. An alternative solution is to learn an optimal stochastic policy, which for each observation of the environment prescribes a probability distribution over available actions that maximizes the agent's average reward per timestep. Such a policy achieves the best performance possible when the agent acts based on immediate observations alone. A reinforcement learning algorithm which learns locally optimal stochastic policies was proposed by Jaakkola, Singh and Jordan, but not experimentally verified. In this talk I will present this algorithm, discuss its implementation, and demonstrate its effectiveness in solving several POMDPs including a maze, a matrix game, and a dynamic routing problem.