Place, U. T. (1993i). Following 'the natural lines of fracture': Concept formation in neural networks [Conference presentation, presented at the Symposium on Associationism, Behaviour Analysis and Connectionism, held at the Annual Conference of the Experimental Analysis of Behaviour Group, University College, London 31st March 1993].
[Abstract]It is an implication of Darwin's theory of evolution by variation and natural selection that the survival and reproduction of complex free-moving living organisms, animals in other words, depends on their ability to change the spatial relations between themselves and other objects, including other organisms of the same and of different species, and so bring about the conditions necessary for that survival and reproduction. In order to do that the organism requires a system - its nervous system - whose function is to match the output to the current stimulus input on the one hand and the organism's current state of deprivation with respect to conditions required for its survival and successful reproduction on the other. Matching behaviour to the conditions required for survival and reproduction is the function of the motivational/emotional part of the system. Matching behaviour to current stimulus input is the function of the sensory/cognitive part of the system. The sensory/cognitive system cannot perform its function successfully without the ability to group inputs together in such a way that every actual and possible member of the class or category so formed is a reliable indicator of the presence of an environmental situation in which a particular behavioural strategy or set of such strategies is going to succeed. In other words the survival and reproduction of an organism of this kind depends crucially on its having a conceptual scheme, a conceptual scheme moreover, which reliably predicts the actual behaviour-consequence relations operating in the organism's environment. Although verbs such as ‘classifying’, ‘categorizing’ and ‘conceptualizing’ are not to be found in Skinner's writings, there is an important passage in The Behavior of Organisms (Skinner 1938) where he addresses the issue which others talk about when they use such terms. Thus in Chapter One, after outlining his "System of Behavior", he goes on to say The preceding system is based upon the assumption that both behavior and environment may be broken into parts which retain their identity throughout an experiment and undergo orderly changes. If this assumption were not in some sense justified, a science of behavior would be impossible. But the analysis of behavior is not an act of arbitrary sub-dividing.  We cannot define the concepts of stimulus and response quite as simply as ‘parts of behavior and environment’ without taking account of the natural lines of fracture along which behavior and environment actually break. (Skinner 1938 p.33). What Skinner has primarily in mind in this passage is the way the scientist's concepts need to be shaped into conformity with what he calls "the natural lines of fracture." But on the Darwinian argument the same must be true of the stimulus classes within which any living organism's behaviour generalises and between which it discriminates. It is argued that studying the properties of artificially constructed neural networks helps us to understand how the brain develops patterns of generalisation and discrimination which do indeed "follow the natural lines of fracture along which behavior and environment actually break." Attention is drawn to the role of the ‘hidden layer’ in responding to resemblances of pattern, to the role of re-entrant/recurrent and reverberatory circuits in establishing expectations on the basis of consecutive stimulus patterns, and to the role of error-correction in bringing stimulus classes into line with the contingencies experienced during learning.
[References]  [Talks]  
Download: 1993i Following 'The Natural Lines of Fracture' - Concept Formation in Neural Networks.pdf