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

An Analytically Tractable Model of Optimal Schema Learning and Few-Shot Generalization

Valentina Njaradi¹, Rachel A Swanson¹, Clémentine Carla Juliette Dominé², James E Fitzgerald³, Andrew M Saxe¹; ¹University College London, ²Institute of Science and Technology, ³Northwestern University

Presenter: Valentina Njaradi

Adaptive behavior depends on the ability to learn rapidly and generalize across new environments — a process supported by the gradual acquisition of reusable schemas over the lifespan. We present a tractable theory of how extended experience enables extraction of low-dimensional structure that supports few-shot generalization. Schema extraction is formalized as an autoencoder that captures environmental regularities, and adaptation to new tasks is modeled by training a downstream readout on the learned latent space. In the high-dimensional regime, we derive exact analytical expressions for generalization and memorization performance as a function of schema dimensionality, amount of data, and task parameters. We find optimal schemas take different forms depending on the learning goal: optimizing for memorization favors high-dimensional representations that preserve episodic detail but generalize poorly, while optimizing for generalization yields compressed, lower-dimensional schemas that support rapid new learning. Schemas optimized for generalization, but not memorization, reproduce key empirical findings from psychology and neurobiology, including accelerated learning under strong schemas, schema-induced distortions, and rapid cortical integration of new memories in the presence of robust prior schemas. Together, these results demonstrate how complementary learning systems can transform naïve representations into expert-like structures that support flexible adaptation.

Topic Area: Memory, Learning & Knowledge Structures