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

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Dynamics and generalization in kernel models of category learning

Matteo Alleman¹, Samuel Lippl¹; ¹Columbia University

Presenter: Matteo Alleman

We study implicit category learning through the lens of kernel models, which are similarity-based models grounded in specific representations. In the setting of binary features, we analytically characterize both the learning dynamics and generalization of these models. In the 3-bit case, we identify three qualitative regimes of learning dynamics, and show that generalization reduces to nearest-neighbour copying, with ties broken by the eigenvalue structure of the kernel. These results hold for SVMs and lazy deep networks trained on cross-entropy, beyond the MSE setting of our theory. Our framework thus connects representational geometry directly to behavior, across both the time-course of learning and the pattern of generalization.

Topic Area: Memory, Learning & Knowledge Structures