Paper: Are Pathology Foundation Model Latent Representations Invariant to Rotation?

iScienceLuvr · x · 2026-08-25

This study investigates the rotational invariance of latent representations across 12 foundation models for digital pathology. By quantifying alignment between non-rotated and rotated patches, models trained with rotation augmentation showed significantly greater invariance. The findings suggest that the lack of rotational inductive bias in transformers necessitates such augmentation during training.

Original post →

More from Research

Research channel →