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

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

Resolving the tension between exemplar and structure learning through a mixture-of-experts model of hippocampal-PFC interactions

Dhairyya Singh¹, Ashley B Williams, Anna C Schapiro¹; ¹University of Pennsylvania, University of Pennsylvania

Presenter: Dhairyya Singh

The hippocampus has been strongly implicated in both taking snapshots of individual experiences and extracting common structure across these experiences—functions often in tension. Prior evidence suggests an anatomical division of labor: the trisynaptic pathway (TSP) employs pattern-separated representations that store episodes while the monosynaptic pathway (MSP) uses overlapping representations to support statistical learning. A fundamental mystery remains, however: how does the brain recruit the right representation at the right time? Prefrontal cortex (PFC) has been proposed to exert control over hippocampal outputs. Here, we introduce a stimulus-computable mixture-of-experts system featuring MSP- and TSP-like neural network experts, along with a PFC-inspired gating network that controls their outputs. The system performs exemplar recognition and categorization simultaneously, and learns to adaptively combine expert outputs. We found that joint training of the experts and the gating network is necessary and simple mixing is insufficient. This framework illustrates how PFC control may harness hippocampal specialization to resolve opposing computational objectives.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF