Place, U. T. (1991d). The problem of error-correction in connectionist networks: a new perspective on the law of effect [Conference presentation abstract]. Proceedings of 17th Annual Convention of the Association for Behavior Analysis - May 24-27 1991 - Atlanta, Georgia (p. 148). Society for the Advancement of Behavior Analysis.
[Abstract]The parallel distributed processor (PDP) derives its ability to learn complex and subtle pattern discriminations from changes that occur in the so-called 'weights' of the synaptic connections between the multiple interconnected nodes of which the device consists as a consequence of previous activation of those connections. In order for the device to have that property, the changes which occur in the synaptic weights must conform to what is known as 'a learning rule' or to a limited number of such rules depending on the circumstances. McClelland and Rumelhart (1988) propose two such rules: "the so-called Hebbian or correlational learning rule .... and the error-correcting or 'delta' learning rule." Attention is drawn to the analogy, if not coincidence, between these learning rules and such traditional learning principles as association by contiguity in the case of the Hebbian or correlational learning rule and the Law of Effect in the case of the error-correcting or 'delta' rule. It is suggested that both of these learning rules are needed, the correlational/contiguity rule to control input selection by the mechanism of selective attention and the 'delta' rule/Law of Effect to control output selection. It is argued that, in the case of living organisms, a second input selection principle is required in order to account for the preoccupation of attention with the motivationally significant at the expense of the motivationally insignificant. Identification of the error-correction rule with the Law of Effect draws attention (a) to the fact that in living organisms the error- and correct-messages are constituted by the immediate consequences of behavior, and that these are differentiated into error-messages (disinforcements) and correct-messages (reinforcements) by the motivational attitude of the organism to those consequences, (b) to the fact that, in contrast to the prejudices of learning theorists in the nineteen thirties who were inclined to deny the weakening effect of punishment on the strength of response tendencies, on the connectionist view it is error-correction (disinforcement) rather than confirmation (reinforcement) that does all the work. It is concluded that what we need is a symmetrical conception of the Law of Effect which allows both the strengthening of response tendencies by success and the weakening of response tendencies by failure.
[References] [Related] [Talks]
Place, U. T. (1991h). Error-correction in connectionist networks: A new perspective on the law of effect [Unpublished paper. Presented to the Annual Conference of the British Psychological Society, Bournemouth, 12th April 1991, Session on Behavioristic Perspectives on Cognitive Psychology and to the 17th Annual Convention of the Association for Behavior Analysis, Atlanta, Georgia, May 26th 1991.] .
[References] [Related] [Talks]
Download: 1991h Error Correction in Connectionist Networks - A New Perspective on the Law of Effect.pdf