Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Neural Mechanism of Compositional Generalization

Zilu Liang¹, Miriam Klein-Flugge, Christopher Summerfield¹; ¹University of Oxford

Presenter: Zilu Liang

Compositionality is a fundamental feature of cognition. Humans can break down learned knowledge into constituents and reassemble them flexibly to solve new problems. Using fMRI, we investigated the neural underpinnings of compositionality in a task where discrete features indicated locations on spatial axes. Participants were trained on a subset of stimuli with feedback and tested on the held-out without feedback. Successful generalization required decomposing the trained stimulus-location associations into rules and recombining the rules to solve the test. Different brain regions adopted distinct representation strategies: rules were represented in high-dimensional parallel manifolds in the hippocampus and composed by vmPFC to solve the test; the superior parietal gyrus put the test stimuli into a low-dimensional spatial reference frame; V1 represented both training and test stimuli as their locations on the ground-truth map.

Topic Area: Predictive Processing & Cognitive Control

Extended Abstract: Full Text PDF