Latent Variable and Implicit Models for Neural System Identification

Abstract

One of the major goals of neural system identification is to understand the underlying neural mechanisms that give rise to visual perception and sensation. While the quest for understanding visual perception goes back many centuries, with the technological advancements in the past decades, machine learning methods have been increasingly used to analyze and model neural responses recorded from various visual sensory areas. In particular, deep neural networks (DNNs) have achieved state-of-the-art performance in predicting the activity of neurons in these regions. These networks have also been shown to learn representations of stimuli that closely match those found in the brain.

Besides their utility as hypotheses about the functional and structural properties of the brain, these powerful predictive models of sensory neurons, sometimes called \textit{digital twins}, allow us to conduct experiments that are not feasible to conduct with their biological counterpart. Importantly, the findings of the experiments conducted with these digital twins have been verified in-vivo, providing further evidence that these models do indeed capture the complex functional properties of visual sensory neurons.

In this thesis, I will discuss three projects that leverage recent advancements in using DNNs to model the responses of visual sensory neurons. The first project focuses on a hybrid model that combines DNNs with latent variable models. This model aims to accurately predict the distribution of neural responses to unseen stimuli. It also infers latent state structures that have meaningful relations to behavioral variables, such as pupil dilation, as well as to the functional and anatomical properties of visual sensory neurons. The second project discusses a model that learns a reparameterization of the stimulus and, combined with DNN-based predictive models, learns a manifold in the stimulus space that visual sensory neurons are equally and maximally responsive to. The third and final project addresses an essential aspect of model development: finding better ways to quantify how good these models are in capturing neural responses.

Overall, this thesis focuses on three important aspects of neural system identification: (1) developing models that account for multiple driving factors of neural responses, (2) showcasing how these models can be used to generate insight into functional and structural properties of visual sensory neurons, and (3) developing metrics that assess the quality of such models.