The estimation and explanation of tuning curves in intermediate visual areas

Abstract

In sensory neuroscience, the characterization of neuronal tuning curves is the de facto method for understanding the bulk of cortical sensory representation. In this thesis, I make two primary contributions to this approach. First, I provide validated statistical estimators that account for the corrupting effect of trial-to-trial variability on the correlation between neural tuning curves and between a noiseless model and a tuning curve. Second, I develop a single-neuron approach to fitting visual cortical responses to deep neural network (DNN) models and demonstrate a method to attribute tuning curve properties of single units in a DNN to the covariance structure of their inputs, which is key to understanding selectivity and invariance. Taken together these methods support the accurate quantification of fundamental summary statistics of noisy neural tuning curves and fine-grained characterization by DNN models of sensory tuning curves. I discuss how this can support the overarching goal of understanding biological intelligence.