Deep Learning Models for Neural Encoding in the Early Visual System

Abstract

One of the questions at the heart of neuroscience is how external experiences are translated into the language of neural spiking. Deemed the neural encoding problem, the relationship between a given stimulus and the resulting neural activity has long been probed by classical statistical methods. In this work, we seek to apply deep learning, an increasingly well-known and powerful way of modeling artificial neural networks, to study this problem in the early visual system. Deep learning is just beginning to be applied in such a way. We take it further by embracing a broad range of network architectures in order to highlight the ways in which the fundamental content of a neural signal changes from when it leaves the retina to when it reaches the brain, and how that content is molded by the different cells at each location. In attempting to model the fundamental way in which neurons communicate–their spike rate–we test our understanding of the underlying principles governing neural coding, explore benefits and drawbacks of past approaches, and achieve new levels of success at matching neural outputs.