Selective networks and recognition automata.

Abstract

The results we have presented demonstrate that a network based on a selective principle can function in the absence of forced learning or an a priori program to give recognition, classification, generalization, and association. While Darwin II is not a model of any actual nervous system, it does set out to solve one of the same problems that evolution had to solve--the need to form categories in a bottom-up manner from information in the environment, without incorporating the assumptions of any particular observer. The key features of the model that make this possible are (1) Darwin II incorporates selective networks whose initial specificities enable them to respond without instruction to unfamiliar stimuli; (2) degeneracy provides multiple possibilities of response to any one stimulus, at the same time providing functional redundancy against component failure; (3) the output of Darwin II is a pattern of response, making use of the simultaneous responses of multiple degenerate groups to avoid the need for very high specificity and the combinatorial disaster that would imply; (4) reentry within individual networks vitiates the limitations described by Minsky and Papert for a class of perceptual automata lacking such connections; and (5) reentry between intercommunicating networks with different functions gives rise to new functions, such as association, that either one alone could not display. The two kinds of network are roughly analogous to the two kinds of category formation that people use: Darwin, corresponding to the exemplar description of categories, and Wallace, corresponding to the probabilistic matching description of categories. These principles lead to a new class of pattern-recognizing machine of which Darwin II is just an example. There are a number of obvious extensions to this work that we are pursuing. These include giving Darwin II the capability to deal with stimuli that are in motion, an ability that probably precedes the ability of biological organisms to deal with stationary stimuli, giving it the capability to deal with multiple stimulus objects through some form of attentional mechanism, and giving it a means to respond directly and to receive feedback from the world so that it can learn conventionally. Already, however, we have shown that a working pattern-recognition automaton can be built based on a selective principle. This development promises ultimately to show us how to build recognizing machines without programs and to provide a sound basis for the study of both natural and artificial intelligence.