Deep Neural Network-Based Simulations for Predicting, Controlling, and Characterizing BOLD fMRI Responses
Loading...
Authors
Shinkle, Matthew
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
activation maximization , brain simulation , encoding models , fMRI , interpretability , neuroscience
Alternative Title
Abstract
Through the work presented in this dissertation, I explore how deep neural networks (DNNs) can be used to simulate, modulate, and characterize BOLD fMRI responses. In my first set of experiments, I developed a set of techniques that transform pretrained DNNs into predictive models capable of simulating BOLD responses to diverse visual stimuli. Comparisons of these simulated responses to real BOLD data show that this method accurately predicts responses in many visual cortical areas. I tested the validity of simulated responses on both naturalistic and highly controlled stimulus sets, showing that simulated data can be used to characterize real selectivity in the brain. In my second set of experiments, I explored how a DNN-based visualization technique called activation maximization can be applied to these simulation models to produce images optimized for specific cortical regions. Analysis of real BOLD responses to these images shows robust modulation of BOLD responses in targeted regions, both within and across participants. I then attempted to push these techniques to their limits, revealing both strengths and challenges of this approach. Finally, in a third set of experiments, I applied my DNN-based simulation techniques to investigate selectivity to naturalistic egomotion in scene-selective cortical areas. Analyses of simulated responses to naturalistic and hand-crafted video stimuli highlight key differences in selectivity between these regions, contributing to our understanding of their distinct functional roles. Taken together, these experiments demonstrate both the utility—and limitations—of DNN-based simulations of BOLD responses. This work serves as an important advancement of the intersection between computational neuroscience and artificial intelligence.