Study decomposes 162 vision models to find universal representations aligned with biological vision

burny_tech · x · 2026-09-27

The paper decomposes the object similarity structure of 162 diverse vision models into a small set of non-negative dimensions, then measures how often each dimension recurs across models.
Universal dimensions are more interpretable and driven by conceptual image properties; architecture, objective, training data, model size, and performance do not explain their emergence.
Models with more universal dimensions better predict macaque IT activity and human similarity judgments, suggesting universality reflects representations relevant to biological vision.
Implication: interpretability and semantic content are implicit factors driving representational convergence across models.

More from Multimodal