A Gradient Descent Approach to Large-Scale Shared Response Modeling for fMRI Analysis

Abstract

As neuroscience research continues to expand its use of machine learning as a critical tool for understanding the functions of the human brain, it becomes increasingly important to generate computationally efficient solutions for working with imaging data. We specifically choose to focus on efficiencies with multi-dataset, multi-subject functional magnetic resonance imaging (fMRI) data, because multi-subject fMRI data lends itself best to making generalizable claims about human cognition. The shared response model (SRM) is one effective framework that analyzes fMRI data of subjects exposed to various stimuli. It is a method that aligns fMRI responses from different brains in a common, information space. SRM works especially well for small datasets. However, with large datasets that have multiple stimuli, multiple runs, and shared participants, fitting SRM becomes difficult because of the high amounts of memory and computational power needed.

In previous work, it was identified that the bottleneck of computational efficiency in SRM is with a specific calculation performed: singular value decomposition (SVD). In this work, we implement and test a mathematical methodology that works around this bottleneck computation in the current implementation of SRM. The aim of this thesis is to replace the SVD computation with an alternative approach that uses optimization by balance-regularized gradient descent. We will perform a number of experiments evaluating the models we propose against the SRM benchmark model. We ultimately conclude that the two models we present have high scalability as compared to the benchmark model while maintaining competitive accuracy.