Adequacy of commonly used similarity measures for capturing computational properties

Ascertain whether commonly used representational similarity measures—including linear regression, Centered Kernel Alignment (CKA), angular Procrustes distance, and Normalized Bures Similarity (NBS)—adequately represent the computational properties of interest when comparing artificial models and neural datasets, and characterize the limitations of these measures.

Background

The paper highlights that, despite widespread use, it is uncertain whether popular similarity measures actually capture the computational properties researchers care about when aligning models with neural data. This uncertainty motivates a systematic evaluation of their limitations.

Establishing the adequacy and limitations of these measures directly impacts claims about brain-likeness of models and the interpretation of cross-system representational alignment.

References

However, while these measures are actively used and provide an efficient way to compare structure across complex systems, it is not clear that they adequately represent the computational properties of interest, and there is a need to better understand their limitations.

— Differentiable Optimization of Similarity Scores Between Models and Brains  (2407.07059 - Cloos et al., 2024) in Introduction

That is why the choice among these measures is still open after a decade.

— A Better Metric Will Not Save You: What You Compare Beats How You Compare It  (2610.13476 - Khaous et al., 9 Oct 2026) in Section 1, Introduction