On Parallel Networks for Optimum Classification.

In this memorandum we study the parallel implementation of optimum classifiers. We first present a parallel implementation of the optimum (or maximum likelihood Gaussian) classifier that finds the output vector with minimum Euclidean distance from the input vector very rapidly. This implementation has the feature of easily cascadable chips allowing the number of output vectors to easily grow to arbitrary size. This network can be incorporated in systems that recognize patterns in time-dependent signals and serve as a shift-invariant associative memory. We then show how to modify the network so that it outputs likelihoods rather than the maximum likelihood vector, and permits different a priori probabilities for the output vectors.