This paper describes a biological approach to automaton theory. It presents a qualitative set of concepts that are checked by simulation studies. The paper differentiates between the biological and computer system approach by the need for the problem to be defined in the case of computer models. (Categories are formulated and stimuli categorized.) Pattern recognizers are based on selection processes in biological models using two procedures, viz, comparison with exemplars and probabilistic feature matching.
Taking analogues with the brain, the authors argue that the basis of neuronal group selection theory is based on the three criteria set out below:
(1) a collection of variant entities (repertoire) capable of responding to the environment,
(2) sufficient opportunities for the entities to encounter the environment, and
(3) a mechanism to enhance/amplify differentially the numbers/strengths of those entities whose responses to the environment are in some sense adaptive.
There must be a balance of the specificity of the recognition. This is covered in neural nets by the concept of degeneracy that gives some redundancy in recognition. The authors set up a simulation model of two recognizers, corresponding to the exemplar and probabilistic categorization. In the first case, the recognizer responds to local features and another transforms the features into an abstract transform. In the second approach, the trace is developed and virtual groups are formed and resolved by the abstracting network. The second level recognizers are corrected and resolved.
The results of the simulation give good correlation with expected results and show the useful analogy with the human system.
