Adaptive Efficient Coding: A Variational Auto-encoder Approach

Abstract

We study a model of neural coding with the structure of a variational auto-encoder. The model posits that the encoding of individual stimulus values is optimally adjusted for a finite training sample of stimuli retained in memory. We demonstrate that this model can rationalize existing experimental evidence on both perceptual discrimination thresholds and neural tuning curve widths in multiple sensory domains. Finally, since our model implies that encoding is optimized for a sample from the environment, it also provides predictions about the adaptation of neural coding as the environmental frequency distribution changes.