Conjectures on mammalian neuron networks for visual pattern recognition

Abstract

Based on Hubel's findings in the visual cortex of the cat, a short-line-extractor neuron configuration is considered as a correlative feature-extraction unit for visual pattern recognition. An array of 19 of these feature-extraction neurons is applied to test patterns consisting of the ten decimal digits. Subsequent processing by a group of memory neurons is equivalent to matrix multiplication. When all of the memory neurons are clamped to the same maximum output level, we get a symmetrical feature-difference matrix whose entries correspond to the summation of feature differences between incoming and memory patterns.