Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Representation can dissociate from function and behaviour in task-optimised vision networks

Marvin Theiss¹, Ben Fausten², Felix A. Wichmann³; ¹University Tübingen, ²Eberhard-Karls-Universität Tübingen, ³University of Tübingen

Presenter: Marvin Theiss

Building models that explain how biological brains process information is a central goal of computational neuroscience. A common way to evaluate such models is to ask how well they align with neural data. This alignment is routinely quantified by comparing model and brain representations using measures such as RSA and CKA. In doing so, high representational similarity is often taken as evidence of similar computation and behaviour. Here, we show that representational similarity can dissociate from both function and behaviour in task-optimised vision networks. Across architectures and datasets, including comparisons to macaque IT cortex, networks can match a target model’s outputs, task performance, and trial-by-trial behaviour while exhibiting a wide range of representational similarity scores. Conversely, networks can closely match task performance and representational similarity, yet differ substantially in trial-by-trial behaviour. These findings show that representational similarity alone is insufficient evidence for shared computation and motivate stricter criteria for brain-model alignment.

Topic Area: Methods, Tools, Theory & Neural Coding