Neural likelihood

Abstract

A large body of evidence shows that perceptual decision making in humans and animals accounts for uncertainty in the relevant stimulus variable. This suggests that the decision is based on a distribution over stimuli given the neuronal activity rather than single point estimates. The likelihood over the stimuli captures this uncertainty for a fixed neuronal response. Because the neuronal population response can be high dimensional, estimating a per-trial likelihood can be challenging. Previous work has thus focused on parametric models, which can introduce a bias by ignoring noise correlations. Here, we present a simple yet general method to decode a per-trial likelihood based on neural networks. Our method applies to discrete and continuous, as well as static and time-series data. We demonstrate it on recordings from two experimental visual paradigms in Macaque V1 and V2.