Research map / Reinforcement learning / Visual and representation RL

A Multi-Region Brain Model to Elucidate the Role of Hippocampus in Spatially Embedded Decision-Making

ICML 2025 · Reinforcement learning › Visual and representation RL · approach: Inferring decision-making from behavior · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds a multi-region brain model (Vector-HaSH+ grid/place cell circuit plus cortical RNN policy trained with REINFORCE) and counterfactually tests grid-cell coding and information flow on the accumulating tower task, finding joint position-evidence grid coding learns most efficiently and reproduces hippocampal conjunctive place fields.

Problem

How the entorhinal-hippocampal-neocortical circuit contributes to spatially embedded evidence-accumulation decisions, and why hippocampal place cells conjunctively encode position and accumulated evidence even when space is task-irrelevant.

Approach

Extends the Vector-HaSH hippocampal-entorhinal scaffold (Vector-HaSH+) with a cortical action-selection RNN trained by policy gradient, and compares counterfactual variants in which grid cells encode position and evidence jointly versus disjointly, with and without sensory input to the hippocampus.

Key results

• Joint grid-cell models M3 and M5 learn more efficiently than disjoint M4 and position-only models; M1 and M2 reach about 47-49% success at LR 1e-4 to 5e-4 while M5 reaches 97-99%.
• Only joint integration models (M3, M5) reproduce choice-specific place cell sequences and conjunctive E x Y place fields reported by Nieh et al. (2021).
• M5 is the only variant whose first PCs of hippocampal activity separate position and local evidence velocity; sensory input to HPC increases exploration efficiency (fewer steps per episode).

Measured effects reported in the paper

Surprising findings

• Standalone RNN baselines with matched parameter counts (M0, M0+) learn poorly, suggesting EC-HPC structure is critical for temporal integration.
• Accumulated evidence was not separable in the first three hippocampal PCs despite grid cells encoding it.

Limitations

• Hippocampal input to cortex is modeled as the only cortical input; MEC and LEC pathways to cortex are not evaluated.
• Grid coding scheme is hand-designed to test joint versus disjoint coding rather than learned.
• Results are from simulations; predictions require future neurophysiological experiments; CA3 recurrence only tested in M2 and M4.

Experimental setup

Datasets: simulated accumulating tower task environment

Benchmarks: accumulating tower task (simulated)

Metrics: cumulative success rate exploration time (steps per episode) mutual information of place cells PCA separability

Baselines: M0 (RNN with sensory input) M0+ (RNN with sensory, positional and evidence velocity) M1 (position-only grid cells) M2 (position-only grid cells, mixed hippocampal input) M3 (joint grid cells, no sensory input to HPC) M4 (disjoint grid cells, mixed HPC input)

Builds on: Vector-HaSH REINFORCE Tolman-Eichenbaum Machine

Closest papers by meaning

• REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 2025 · similarity 0.87)
• Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.87)
• A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.87)
• Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps (NeurIPS 2024 · similarity 0.87)
• Not so griddy: Internal representations of RNNs path integrating more than one agent (NeurIPS 2024 · similarity 0.86)
• Leveraging Attractor Dynamics in Spatial Navigation for Better Language Parsing (ICML 2024 · similarity 0.85)