Systolic Pattern Recognition Based on Neural Network Algorithm

Abstract

The paper presents a solution for pattern classification, which uses distribute processing both for computing the matching score and selecting the class with the maximum score. The proposed architecture belongs to systolic arrays, being a generalization of the classical priority queue. A detailed description of the elementary processors (EPs) reveals that the algorithm implemented by each EP (which is based on computing the Hamming distance) is common also for neural networks. The overall result is a O(M) execution time for M classes (i.e. linear), and O(1) execution time with respect to n (the size of the patterns).

For testing the ideas, a simulator has been developed. It has been built starting from a set of C functions for simulating parallel processes. A short description of these functions supports our claim about the improvement of efficiency when developing a simulator starting from these functions. Several results are shortly discussed. Conclusions and further directions of research end the paper.