Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

An Automated Approach Using Sparse Dictionary Learning to Discover Visual Representations in the Human Visual Cortex

Qi, Qinglin LU (2026) PSYP01 20261
Department of Psychology
Abstract
Visual features are encoded by the population activity of neurons across the human visual hierarchy, from primary visual features in early areas to semantic visual features in the higher visual cortex. Sparse dictionary learning and sparse autoencoders have recently emerged as powerful automated and unsupervised methods for discovering monosemantic representations of features in the massive data of artificial and biological neural networks. The present study set out to discover, validate, and interpret the visual features encoded by fMRI activity across the visual hierarchy by analyzing a publicly released dataset of 7T fMRI responses collected from eight participants viewing over 70,000 natural images, with additional data augmentation.... (More)
Visual features are encoded by the population activity of neurons across the human visual hierarchy, from primary visual features in early areas to semantic visual features in the higher visual cortex. Sparse dictionary learning and sparse autoencoders have recently emerged as powerful automated and unsupervised methods for discovering monosemantic representations of features in the massive data of artificial and biological neural networks. The present study set out to discover, validate, and interpret the visual features encoded by fMRI activity across the visual hierarchy by analyzing a publicly released dataset of 7T fMRI responses collected from eight participants viewing over 70,000 natural images, with additional data augmentation. The study focused particularly on expanding the scope of visual feature discovery, separating feature recovery from its interpretation, and reducing agentic bias in discovering and interpreting visual representations. Participant-wise fMRI responses were decomposed by a sparse autoencoder into localised fMRI activity patterns. Each pattern was independently evaluated under two dedicated streams. In total, 35,181 and 6,603 fMRI patterns were discovered, validated, and interpreted, representing 113,714 semantic and 6,870 primary visual features, respectively. Two limitations identified in past literature were also tested quantitatively. The findings confirmed the dissociation between the selectivity and the interpretability of primary visual features, and that interpreting primary visual features as semantic risks representational misattribution. The present work contributes a novel method for the automated discovery of visual representations in human fMRI data, emphasising the authenticity of representation discovery and interpretation. (Less)
Please use this url to cite or link to this publication:
author
Qi, Qinglin LU
supervisor
organization
course
PSYP01 20261
year
type
H2 - Master's Degree (Two Years)
subject
keywords
human vision, visual cortex, visual representation, functional magnetic resonance imaging, sparse dictionary learning
language
English
id
9251074
date added to LUP
2026-09-15 10:56:52
date last changed
2026-09-15 10:56:52
@misc{9251074,
  abstract     = {{Visual features are encoded by the population activity of neurons across the human visual hierarchy, from primary visual features in early areas to semantic visual features in the higher visual cortex. Sparse dictionary learning and sparse autoencoders have recently emerged as powerful automated and unsupervised methods for discovering monosemantic representations of features in the massive data of artificial and biological neural networks. The present study set out to discover, validate, and interpret the visual features encoded by fMRI activity across the visual hierarchy by analyzing a publicly released dataset of 7T fMRI responses collected from eight participants viewing over 70,000 natural images, with additional data augmentation. The study focused particularly on expanding the scope of visual feature discovery, separating feature recovery from its interpretation, and reducing agentic bias in discovering and interpreting visual representations. Participant-wise fMRI responses were decomposed by a sparse autoencoder into localised fMRI activity patterns. Each pattern was independently evaluated under two dedicated streams. In total, 35,181 and 6,603 fMRI patterns were discovered, validated, and interpreted, representing 113,714 semantic and 6,870 primary visual features, respectively. Two limitations identified in past literature were also tested quantitatively. The findings confirmed the dissociation between the selectivity and the interpretability of primary visual features, and that interpreting primary visual features as semantic risks representational misattribution. The present work contributes a novel method for the automated discovery of visual representations in human fMRI data, emphasising the authenticity of representation discovery and interpretation.}},
  author       = {{Qi, Qinglin}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{An Automated Approach Using Sparse Dictionary Learning to Discover Visual Representations in the Human Visual Cortex}},
  year         = {{2026}},
}