ANALOG COMPUTATIONAL MODELS OF CONCEPT-FORMATION

This paper proposes and describes a method of inductive concept learning, a method suitable for implementation in parallel computational mode with analog VLSI neural-net circuitry. The approach is consonant with the original Perceptron approach. However, weights along linear links are not learned adaptively. Instead, the net depends upon the frequency of occurrence to adjust the strength of activation generated by an input and the attention paid to the input. Of critical importance are the relative magnitudes of the information complexity of the concept to be learned and the complexity of the implementation hardware. If the former exceeds the latter, the concept cannot be learned. The manner in which failure is signaled and hardware complexity is increased is described in this paper.