A neural geometry theory comprehensively explains apparently conflicting models of visual perceptual learning

Abstract

It is established that perceptual learning enhances the signal-to-noise ratios of sensory processing. However, the specific mechanisms behind this improvement remain enigmatic. Here, we systematically investigated these mechanisms by analyzing population activity changes in deep convolutional neural networks (DCNNs), humans, and macaques. We first trained DCNNs on orientation and motion direction discrimination tasks. Our models successfully replicated the behavioral signature of enhanced signal-to-noise ratio, and the individual and population response changes observed in numerous empirical studies. We then devised a method to quantify all scenarios in which learning modifies the tuning and covariance structure of neuronal populations to enhance sensory discrimination. A groundbreaking revelation is that enhanced signal-to-noise ratios are primarily driven by reduced trial-by-trial response variability. This finding was strongly supported by population response changes measured in visual areas of both humans and macaques. These results elucidate the shared mechanisms of neural plasticity in biological and artificial visual systems.