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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Universal and idiosyncratic dimensions in vision and language models

Johannes Singer¹, Alessandro Thomas Gifford¹, Lingcong Kong¹, Radoslaw Martin Cichy¹, Adrien Doerig¹; ¹Freie Universität Berlin

Presenter: Johannes Singer

Recent findings suggest that diverse artificial neural network models (ANNs) converge on a set of "universal" brain-aligned representational dimensions despite variation in architecture, task, and training data. This raises doubts about how much insight can still be gained from comparing ANN models in terms of their brain alignment. Here, we systematically assess similarities and differences between representational dimensions in vision and language ANNs. We find a subset of strongly brain-aligned dimensions that are universal across models, consistent with a convergence of diverse models. However, we also find systematic differences: even highly brain-aligned dimensions show a gap in cross-model generalization, map to distinct brain regions, and encode interpretable information. These results show that, alongside universal representational dimensions, idiosyncratic dimensions capture complementary aspects of neural representations. This highlights that idiosyncratic representational dimensions can be a critical source of insight and should not be overlooked when using ANNs to study neural representations.

Topic Area: Methods, Tools, Theory & Neural Coding