Date of Award

Fall 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Psychology

First Advisor

Chun, Marvin

Abstract

Convolutional neural networks (CNNs) are the state-of-the-art artificial vision models for object recognition tasks. Their architecture is inspired by the organization of the ventral visual system of the brain, and although the way they are trained is not constrained by neural responses, the activations of artificial neural networks nevertheless show some similarities to activations of primate visual regions. Here, across three different studies, I compare the representation of objects in CNNs with that of the primate visual system in two different ways: (1) examining the relationship between object identity and non-identity information and (2) examining the representation of multiple objects. I find that a linear mapping function can successfully link object responses in different states of non-identity transformations in human occipito-temporal cortex (OTC), posterior parietal cortex (PPC), and CNNs for both Euclidean and non-Euclidean features. Overall, object identity and non-identity features are represented in a near-, rather than complete-, orthogonal manner in all three cases. This shows some similarities between CNNs and the primate visual system. Meanwhile, I also find some differences between the two. Although there is multiple object normalization in both individual voxel amplitude response and population pattern response in lateral occipital cortex (LO), CNNs demonstrate a discrepancy between unit and population response with some normalization at the unit level but weak normalization at the population level. This dissertation characterizes new ways in which the primate visual system and CNNs represent objects, and provides insights into how we can improve CNNs to be more brain-like.

Download

Share

COinS