Year: 2023
Type: dissertation
Abstract: In this work, we analyze cross-subject transfer learning (CSTL) for predicting fMRI-measured brain activity in a target subject, using perceptual features of visual processing learned from source subjects. We demonstrate transferability of perceptual feature spaces derived from deep learning networks across different performance metrics. We show perceptual feature spaces outperform baseline and current state-of-the-art functional alignment approaches, such as neural code converters and hyperalignment, at small training sizes for overlapping sets of stimuli between subjects. Additionally, we show inter-individual encoding approaches perform better for lowerorder vs. higher-order visual cortices (LVC and HVC, respectively). Transferability of perceptual feature spaces is more pronounced in the LVC, and less significant in the HVC. Moreover, we show perceptual features generated using latent representations of BOLD signals exhibit better CSTL properties compared to those generated using BOLD signals in voxel space. While shown for overlapping stimuli, this was particularly evident for sets of disjoint stimuli.
Source: TSpace (University of Toronto)
Author: Talukder, Labeeb Shahriar
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7133027403