Place, U. T. (1992j). Towards a reconciliation between the associationist and radical behaviorist traditions in the experimental analysis of behavior. [Unpublished paper. Presented under the title 'The three term contingency as a link between the associationist and radical behaviorist traditions in the experimental analysis of behavior' as Invited Address to the First International Congress on Behaviorism and the Sciences of Behavior, Guadalajara, Mexico, 6th October 1992].
[Abstract]It is an implication of the Law of Non-Contradiction that two incompatible descriptions of the same class of phenomena cannot both be true. This suggests that the future for radical behaviorism must lie in achieving a reconciliation with other disciplines and approaches studying the same or closely related phenomena. The approach known as "associative learning theory" shares a common data basis with radical behaviorism in the area of the experimental analysis of animal behavior. It is separated from radical behaviorism by a different view of the nature of what is learned. According to the radical behaviorist, under certain antecedent conditions (discriminative stimulus + establishing condition) an organism learns to emit a response. According to associative learning theory what is learned is an association between a pair of consecutive stimulus events. When presented with the first member of the pair, the organism learns to "predict" or "expect" the second member of the pair. Until recently, the principal application of this principle was Rescorla and Wagner's (1972) analysis of  Pavlovian (respondent) conditioning. More recently, Adams and Dickinson's (1981) reinforcer-devaluation experiment has led associationists to pay more attention to instrumental (operant) learning. It has also opened up an interesting divergence of views between Dickinson (1988; Heyes and Dickinson, 1991; Dickinson & Balleine, forthcoming) who takes it as evidence of a discontinuity between respondent conditioning, which he interprets in terms of the establishment of mechanical associations, and operant learning which he interprets in terms of the ‘beliefs’ and ‘desires’ of philosophical action theory, and Rescorla (1991) who uses it as evidence for an interpretation of operant learning based on the same principles of stimulus-stimulus association invoked by Rescorla and Wagner to account for respondent conditioning. Standing in the way of a reconciliation between radical behaviorism and associative learning theory are the misgivings of the former about the use made by the latter of ‘mentalistic’ concepts, such as ‘expect,’ ‘anticipate,’ and ‘predict.’ These misgivings may be allayed if attention is paid to the results of applying to such concepts the technique, known as ‘conceptual analysis,’ developed by Wittgenstein (1953; 1958) and the philosophers of the Oxford ‘ordinary language’ school. A recent application of this technique to the linguistic phenomenon known variously as ‘intentionality’ or ‘intensionality’ shows that it consists of two distinct varieties of ‘referential anomaly’ which ‘infect’ the grammatical objects of certain verbs. In one case, the grammatical object is used to indicate a range of possible events any one of which, if it were to occur, would constitute a manifestation or satisfaction of a disposition. In the other case, the grammatical object functions as a quotation of what the agent either has said or might be expected to say or have said. Referential anomaly of the dispositional kind is both unavoidable and benign, but the use of quotations to characterize behavioral dispositions is acceptable for scientific purposes only in those cases where the behavior in question is in fact subject to linguistic control. Since the grammatical object of the verbs ‘know,’ ‘believe’ and ‘think,’ as they occur in belief/desire explanations, takes the form of an embedded indicative sentence in oratio obliqua or indirect reported speech, Dickinson's explanation of instrumental/operant learning in animals involves the scientifically unacceptable metaphor of linguistic initiation and control. Rescorla's theory, on the other hand, requires nothing more than that the organism learn to ‘expect’ or ‘anticipate’ an event (the outcome), given the combination of an antecedent discriminative stimulus and the stimulus constituted by the incipient emission of the response which it evokes. In this case the anomaly of reference in the noun phrase which occurs as the grammatical object of the verb reflects its use as a device for indicating a range of possible outcomes any one of which, if it occurred, would fulfill and confirm the expectation which it specifies.
Note:
UTP made changes to the text of the presentation in 1995 and in 1999.
[References]  [Talks]  
Download: 1992j 1999 Towards a Reconciliation between the Ascociationist and Redical Behaviorist Traditions in the Experimental Analysis of Behavior.pdf

Place, U. T. (1993j). Unsupervised and supervised learning in neural networks [Unpublished paper presented at the Annual Conference of the Experimental Analysis of Behaviour Group, University College, London, 30th March 1992, and at the Inter-Univerrsity Centre Conference on 'Connectionism and the Philosophy of Mind,' Park Hotel, Bled, Slovenia, 10th June 1993].
[Abstract]The paper examines the relationship between three distinctions, two drawn from the current literature on learning in connectionist networks and one from the animal learning literature: 1. the distinction drawn by connectionists between 'unsupervised' and 'supervised' learning, 2. the distinction also drawn by connectionists between the Hebbian and 'delta' or error-correction learning rules, and 3. the distinction drawn within traditional learning theory between classical or respondent conditioning on the one hand and instrumental or operant learning on the other. It is argued that, despite differences in the way error-correction is applied in the two cases, the distinction between unsupervised and supervised learning corresponds closely to that between classical and instrumental learning. But, whereas unsupervised learning is usually implemented in artificial networks by a version of the Hebbian rule and supervised learning by the 'delta' rule, recent and not so recent work in animal learning suggests that, given plausible assumptions about the arrangement of the network, a version of the Hebbian rule can account for both types of learning.
[References]  [Talks]