La Gioconda and the indeterminacy of smile recognition by a person and by an artificial neural network

摘要

This paper presents a comparative analysis of the possibilities of smile recognition by a person and by an artificial neural network under conditions of indeterminacy. The main brain-activity patterns are studied by functional magnetic-resonance tomography. There are fundamental limitations inherent to natural and artificial neural networks, and therefore generalizations of the results of recognizing test images are obtained under below-threshold and above-threshold conditions. The probability of recognizing a smile is thus fairly high under ordinary conditions, but it decreases under indeterminacy conditions (threshold and noisy images) both in humans and in artificial neural networks. For instance, the recognition of a smile in La Gioconda's facial expression by a person and by an artificial neural network occurs with probability 0.69. We assume that the most important operating principle in both networks is a matched-filtering mechanism as a measure of how well the presented image corresponds to a pattern learned by the neural network-in particular, a smile. (C) 2020 Optical Society of America