A hybrid cellular automaton/neural network classifier for multi-valued patterns and its VLSI implementation

Abstract

A multi-valued pattern classifier with high discrimination sensitivity and its VLSI implementation proposal on a single chip are presented in this paper. The classification scheme is based on the combination of a reconfigurable Cellular Automaton and a Neural Network architecture. A 2-D Reconfigurable Hybrid Additive Cellular Automaton (RHACA) architecture amplifies the Hamming distance between patterns, whereas a neural network architecture, implemented in digital form, assigns vectors of weighing coefficients which take into account the relative significance of the sites on the 2-D lattice. The proposed classifier is able to operate successfully even for pattern classes of small difference, or for patterns that lie on the decision boundaries between classes. If the training and processing phases are not partitioned, the proposed classification scheme is able to operate in partially exposed environments. With the proper setting of admittance levels into the classes of multi-valued patterns involved, the proposed classifier can also operate on patterns with partly missing data. The proposed multi-valued pattern classifier can be realized on a single VLSI chip with dimensions 7.73 mm × 8.14 mm = 62.96 mm 2 and the expected frequency of operation for the chip is 50 MHz.