Dieu, Patrick-Olivier: Representation Similarity And Transformation Stablility: An Empirical Study of Structural Anti-Correlation in Neural Networks

Representation similarity measures such as Centered Kernel Alignment (CKA) are widely used to compare neural networks by quantifying geometric alignment in representation space. However, their relationship to functional behavior under input transformations remains unclear. We study the relationship between representation similarity and transformation-dependent stability across a sequence of neural networks. We observe a strong and consistent negative correlation between these two quantities, whi