🚀 go-pugleaf

          The University of Western Ontario
            Department of Computer Science

                    COLLOQUIUM


Speaker: Professor Charles Ling, Western Ontario
Title:   Answering Connectionist Challenge
         A symbolic model for learning Past Tenses of English Verbs
Time:    Wednesday, October 14, 1992, at 2:00pm
Place:   MC 320
Abstract:
Are Artificial Neural Networks (ANNs) more adequate models for
``complicated'' human learning tasks than symbolic learning methods?
What kind of tasks can be better learned and generalized
by using ANNs or symbolic methods?
We give some answers to these questions, and study a benchmark task
-- the learning of past tenses of English verbs. Connectionist models
have been claimed to be successful in solving this learning task.
MacWhinney and Leinbach (1991) pose a challenge claiming that no
symbolic methods can perform as well as their connectionist model.
We design a general purpose Symbolic Pattern Associator (SPA)
based on ID3 (see Quinlan (1986)) that can learn and generalize
from any patterns.
In comparison to the connectionist models, the SPA is able to
generalize much better to the past tenses of unseen regular and
irregular verbs. The SPA's performance is also more psychologically
realistic in many aspects. In addition, the SPA is able to represent
the acquired knowledge in production rules which are meaningful to
human observers, and which allow further generalization and integration
into other learning modules. On the other hand, the connectionist models
leave the acquired knowledge in a mysterious black-box.

This represents collaborative work with Steve Cherwen and Marin Marinov.

MacWhinney and Leinbach (1991). Implementations are not conceptualizations:
Revising the verb model. In Cognition, volume 40. Pages 121 - 157. 1991

Quinlan (1986). Induction of decision trees. In Machine Learning,
Volume 1. Pages 81 - 106. 1986.