Reproduce “Universal dimensions of visual representation” on OpenResearch

About this paper

Zirui Chen, Michael F. Bonner

The authors test whether image-response dimensions shared across vision models predict human visual-cortex activity. They measure universality by predicting principal components from other models, and brain similarity by predicting them from fMRI. Across three trained groups, keeping the five or ten most universal dimensions caused little or no decrease in RSA scores, a finding about representational similarity, not computational efficiency.

Key insights

• Brain-aligned representations in deep neural networks are not primarily dependent on specific network architectures or task objectives, but instead reflect universal aspects of natural image representation that emerge convergently across diverse systems.
• The strong alignment observed between universal dimensions and the human brain is a learned property, substantially increasing in trained networks compared to untrained ones.
• A small, core subspace of universal dimensions largely accounts for the representational similarities commonly observed between artificial neural networks and the human visual cortex.

Summary by alphaXiv.