Date of Award

Fall 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Applied Physics

First Advisor

Murray, John

Abstract

The brain is one of the most complex dynamical systems in the universe and despite our everyday familiarity with it, the intricacies of how it functions remain a mystery to us. It is enormously powerful, continuously learning, and able to perform computations flexibly and adapt to novel situations. Although science presently has an understanding of how individual neurons work and a general idea of what the functions of each region in the brain are, how the computations of each neuron are modulated by context and how the various regions work together to produce behavior are unclear. Here, we explore the complex computational mechanisms underlying decision-making in the brain, addressing two fundamental questions: first, how the brain uses information in order to make decisions, and second, how to build models that make decisions more like we do. We approach these fundamental question at both the systems level and at the scale of individual neurons to build up a better mechanistic understanding of how the brain is able to develop strategies for problem solving and how individual neurons modulate their computations in order to give rise to such behavior. For the former question, we examine how context-dependent neuronal activity in four association cortex areas shape the computations performed by individual neurons, revealing that activity is at least as strongly modulated by temporal context as it is by conventional task variables. For the latter question, we introduce a novel hybrid reinforcement learning recurrent neural network model (RLRNN) inspired by the interplay between the prefrontal cortex and the basal ganglia to investigate how the brain deviates from simple reinforcement learning behavior in complex environments. We show that traditional RNN models typically considered analogous to the brain often learn solutions that are not biologically plausible while our RLRNNs are able to capture the nuanced strategic deviations observed in animals, suggesting that the interaction between these two systems in the brain are vital for generating sophisticated and adaptive behaviors.

Download

Share

COinS