Publication Details
Title: Exploiting Temporal Binding to Learn Relational Rules Within a Connectionist Network
Author: L. Shastri
Group: ICSI Technical Reports
Date: June 1997
PDF: http://www.icsi.berkeley.edu/pubs/techreports/tr-91-003.pdf
Overview:
Rules encoded by traditional rule-based systems are brittle and inflexible because it is difficult to specify the precise conditions under which a rule should fire. If the conditions are made too specific a rule does not always fire when it should. If the conditions are made too general, the rule fires even when it should not. In contrast, connectionist networks are considered to be capable of learning soft and robust rules. Work in connectionist learning, however, has focused primarily on classification and feature formation, and the problem of learning rules involving relations and roles (variables) has received relatively little attention. We present a simple demonstration of rule learning involving relations and variable within a connectionist network. The network learns the appropriate correspondence between roles of antecedent and consequent relations as well as the features that role fillers must possess for a rule to be applicable in a given situation. Each rule can be viewed as mapping from the symbolic level to the symbolic level mediated by a semantic filter embedded within a subsymbolic level. The network uses synchronous firing of nodes to express dynamic bindings.
Bibliographic Information:
ICSI Technical Report TR-97-003
Bibliographic Reference:
L. Shastri. Exploiting Temporal Binding to Learn Relational Rules Within a Connectionist Network. ICSI Technical Report TR-97-003, June 1997
Author: L. Shastri
Group: ICSI Technical Reports
Date: June 1997
PDF: http://www.icsi.berkeley.edu/pubs/techreports/tr-91-003.pdf
Overview:
Rules encoded by traditional rule-based systems are brittle and inflexible because it is difficult to specify the precise conditions under which a rule should fire. If the conditions are made too specific a rule does not always fire when it should. If the conditions are made too general, the rule fires even when it should not. In contrast, connectionist networks are considered to be capable of learning soft and robust rules. Work in connectionist learning, however, has focused primarily on classification and feature formation, and the problem of learning rules involving relations and roles (variables) has received relatively little attention. We present a simple demonstration of rule learning involving relations and variable within a connectionist network. The network learns the appropriate correspondence between roles of antecedent and consequent relations as well as the features that role fillers must possess for a rule to be applicable in a given situation. Each rule can be viewed as mapping from the symbolic level to the symbolic level mediated by a semantic filter embedded within a subsymbolic level. The network uses synchronous firing of nodes to express dynamic bindings.
Bibliographic Information:
ICSI Technical Report TR-97-003
Bibliographic Reference:
L. Shastri. Exploiting Temporal Binding to Learn Relational Rules Within a Connectionist Network. ICSI Technical Report TR-97-003, June 1997