In this paper we develop a set of formal metrics for assessing knowledge representation strategies and use them for evaluating the relative strengths and weaknesses of the existing methods. We further propose an alternate knowledge representation schema for neural networks which satisfies most of our criteria. We present an implementation of our approach using a massively parallel architecture for binary image recognition tasks whereby each neuron computes a local signature for the input pattern mapping it onto a structure resembling a torus. The resulting torus has several desirable characteristics including shift invariance property to facilitate efficient recognition of misaligned input. The paper also includes performance analysis of our representation for storing two-dimensional patterns as compared to the existing paradigms for information storage.
