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

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Aligning DNN Representational Spaces with Shepard’s Law

Navya Sahay¹, Daniel L. Carstensen², Steven M Frankland³, Serra E. Favila²; ¹Santa Fe Institute, ²Brown University, ³Dartmouth College

Presenter: Navya Sahay

According to Shepard's (1987) universal law, generalization strength between two stimuli decays as a non-linear function of their distance in psychological space. Recent work suggests that this law also holds in the representational spaces of deep neural networks (DNNs). However, it remains unclear how reliably DNN representations capture Shepard's law or approximate human psychological space, two distinct targets of alignment. We tested whether a lightweight two-layer perceptron head trained to predict human triplet odd-one-out judgments from DNN representations could improve alignment with both targets. Training consistently increased Shepard-like structure in the transformed representations. However, gains in alignment with human psychological spaces depended on the specific DNN and training data. These results suggest that lightweight alignment methods can sharpen Shepard-like structure in DNNs, but their success in capturing human representations depends critically on the choice of model and data.

Topic Area: Methods, Tools, Theory & Neural Coding