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

Simulating hippocampal-mPFC interactions with a mixture-of-experts model

Dhairyya Singh¹, Anna C Schapiro¹; ¹University of Pennsylvania

Presenter: Dhairyya Singh

The hippocampus supports both learning fine-grained details of individual experiences and extracting shared structure across them, despite their competing computational demands. Prior work suggests a division of labor within the hippocampus: the trisynaptic pathway (TSP) builds pattern-separated representations for epi-sodic learning, whereas the monosynaptic pathway (MSP) builds overlapping representations for structure learning. But how does the brain deploy the right path-way for a given task? We introduce a mixture-of-experts framework with MSP- and TSP-like neural network experts and an mPFC-inspired gating network that computes trial-wise weightings of their outputs. Trained end-to-end, the gating network learns to preferentially recruit MSP for categorization and TSP for exemplar learning. We show that the model reproduces findings from rodent contextual fear conditioning experiments, where mPFC disruption impairs fear-memory specificity with-out disrupting fear learning. Adaptive gating shifts recruitment toward TSP when a novel shock context must be encoded, reducing representational overlap with similar contexts. These results offer a mechanistic theory of how mPFC may dynamically engage hippocampal pathways to balance competing memory demands.

Topic Area: Memory, Learning & Knowledge Structures