Characterizing human vision through large-scale brain imaging and computational models

Abstract

Efforts to understand the neural underpinnings of human visual processing require sufficient experimental data and robust models. This thesis significantly contributes to both these fronts while simultaneously elucidating some of the most intriguing aspects of the human visual system. In the first chapter, I use a combination of classical machine learning, artificial neural networks, and a joint MEG/fMRI neuroimaging dataset to reveal that the human visual system extensively processes highly memorable images in regions distributed throughout visual cortex late in time. In the second chapter, I present the BOLD Moments Dataset, a large-scale fMRI dataset using short video stimuli to extend computational models of visual processing into the video domain to better understand how humans process visual content unfolding over time. The last chapter introduces a fMRI dataset aggregation framework titled MOSAIC to achieve the scale and stimulus diversity needed for training modern neural networks directly on brain responses. This body of work exemplifies how large-scale experimental data and artificial neural networks can contribute towards a robust and generalizable understanding of human visual processing.