Back to search

University of British Columbia

Decoding neural representations in higher-order visual cortex : an explainable AI-based fMRI analysis

Abstract

dc:description

Early visual cortex has a retinotopic organization, with neurons tuned to specific properties such as motion and color. However, the governing principles of the higher-order visual cortex remain a topic of ongoing debate. Some studies suggest distinct modular cortical regions for specific image categories, while others propose distributed overlapping representations. This thesis aims to employ deep learning-based functional magnetic resonance imaging (fMRI) analysis combined with explainable artificial intelligence (AI) methods, to explore the topological organization of the ventral temporal visual cortex. Using BOLD5000 dataset with fMRI responses from four participants exposed to thousands of images, we defined eight super-categories: body, face, animal, outdoor, etc. A neural network model was trained for each super-category to discern its presence or absence based on whole-brain fMRI responses. Each model achieved significantly above-chance area under the curve (AUC) performance compared to a randomized model, even when tested on different participants, suggesting a common topological organization underlying visual object representations. Using an explainable AI technique, we identified cortical regions responsible for the model's decisions on each super-category. The fusiform face area (FFA) and occipital face area (OFA) were highly sensitive to faces, while extrastriate body area (EBA) was more sensitive to bodies, and parahippocampal place area (PPA) to outdoor scenes. Other super-categories showed less distinct cortical specialization across participants. To explore further, we assessed the impact of masking specific brain regions. For faces, preserving FFA and OFA maintained the model's AUC, while masking them significantly decreased it but did not affect other super-categories, supporting modular organization. Similarly, masking EBA and PPA lowered performance for body and outdoor super-categories. However, even with masked regions, models still outperformed random models, indicating partial involvement of distributed representations. These findings suggest that both modular and distributed representations coexist, with a stronger modular organization for face, body, and outdoor super-categories. This study leverages advanced explainable AI techniques to bridge the gap between complex deep learning models as neural representation decoders and their interpretability. This methodological framework holds promise for future research in visual neuroscience, paving the way for a richer understanding of the neural basis of visual processing and cognition.

Degree

thesis:*
Name thesis:degree_name
Master of Science - MSc
Level thesis:degree_level
master's
Discipline thesis:degree_discipline
Neuroscience
Grantor dc:publisher
University of British Columbia
Year dc:date
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Khademi, Mahmoud
Contributors dc:contributor
  • Oruc, Ipek

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language
eng

Identifiers

dc:identifier.*
Handle dc:identifier
http://hdl.handle.net/2429/89156

Chain of custody

source
Harvested from
University of British Columbia
Base URL
circle.library.ubc.ca/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Khademi, Mahmoud. Decoding neural representations in higher-order visual cortex : an explainable AI-based fMRI analysis. master's thesis, University of British Columbia, 2024. http://hdl.handle.net/2429/89156