Date of Award

2020

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Graduate Group

Neuroscience

First Advisor

Danielle S. Bassett

Abstract

The human brain supports complex behavior by forming detailed models of statistics underlying real-world experience, including vision, language, and movement, allowing us to correctly make predictions and plan our actions. These statistical models are central to human cognition, and failure to correctly learn statistical models has been implicated in several psychiatric conditions including schizophrenia, autism, major depression, and bipolar disorder. Despite its importance to both healthy cognition and disease, our understanding of how the brain learns statistical models continues to present a significant challenge to the field of cognitive neuroscience.

Recently, network science has shown potential to advance our understanding of how we learn these models. For instance, transforming semantic relationships between English words into a graph allows characterization of common organizational motifs, and presence or absence of these motifs is shown to affect human learning. Further, brain activity itself consists of interactions between functionally diverse cortical regions. Network science is thus well suited to characterize neural mechanisms that support learning statistical models, including functional dynamics that take place during learning and the structural brain connectivity that underlies these dynamics.

Here, we approach statistical model learning through complementary perspectives of human behavior, brain structure, and brain activity. First, we used diffusion-weighted imaging to identify variability in structural connectivity that accounted for individual differences in motor sequence learning. Next, we investigated the role of topology in graph learning, showing that learners exhibit quicker reaction times when sequences are drawn from modular versus lattice graph structure. Finally, we employed functional neuroimaging to identify neural correlates of graph learning, finding systematic differences in stimulus representations when subjects learned from a modular graph versus a ring lattice.

This collection of studies adds to a growing body of knowledge regarding behavioral and neural correlates of graph learning. The studies highlight the crucial role of structural brain connectivity in facilitating sequence learning, and the extent to which graph topology shapes our expectations on simple learning tasks. Our results provide a window into how graph learning could serve to optimize the learnability of informational systems, as well as answer fundamental questions about how humans represent statistical structure.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS