Description
- Abstract:
- This work explores to what extent deep neural networks (DNNs) can serve as accurate and biologically meaningful models of human visual processing. While DNNs have achieved remarkable success in visual recognition benchmarks such as ImageNet, recent findings reveal that increasing model performance on these tasks does not necessarily translate into better alignment with neural activity in the primate brain or human behavioral patterns. The first part of this work revisits the long-held assumption that performance on categorization tasks correlates with neural predictivity. Using benchmark datasets and neural recordings from inferotemporal cortex, we show that this relationship has broken down in modern architectures, which are increasingly less brain-like despite improved accuracy. In response, we propose a harmonization approach that encourages alignment between model visual strategies and those used by humans by constraining the training routine with human behavioral data. Although this strategy improves alignment with neural and behavioral data, it remains biologically implausible. To address this limitation, we develop a new class of models grounded in anatomical constraints and inspired by the classical HMAX framework. These models incorporate architectural features like scale-band processing and feedback, enabling improved generalization to naturalistic images and more robust alignment with neural recordings.
- Notes:
- Thesis (Ph. D.)--Brown University, 2025
Citation
Rodriguez, Ivan Felipe,
"Aligning deep neural networks with biological vision"
(2025).
Cognitive, Linguistic, and Psychological Sciences Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:bg37c9w8/
Relations
Collection:
-
Cognitive, Linguistic, and Psychological Sciences Theses and Dissertations
Theses and Dissertations for the Cognitive, Linguistic, and Psychological Sciences department....