ICML 2026 · Conference paper

The Hippocampal Place Field Gradient: A Bio-inspired Framework Building Multiscale Representation for Better Sample Efficiency

Shujun Zhou, Junrong Qi, Guozhang Chen

Modeling how grid cell projections generate multiscale hippocampal place fields inspires a Transformer positional embedding that improves sample efficiency.

The scientific question

How do discrete grid cell inputs generate continuous multiscale hippocampal place fields?

The hippocampus exhibits a continuous dorsal-ventral gradient of place field sizes, yet its circuit mechanism and computational benefit for learning remained unclear.

Earlier preprint model architecture and context-dependent task design.
Figure 2 from the earlier preprint ↗

The core idea

Frequency-Dependent Grid Projection Decay

A theoretical model derives continuous place field gradients from frequency-dependent decay of grid cell projections, motivating a Hippocampus-Inspired Positional Embedding (HIPE) for Transformer networks.

What we found

Balancing Precision and Sample Efficiency

Multiscale place fields establish an inductive bias that trades off spatial precision against sample efficiency, and incorporating HIPE into Transformers improves data-efficient generalization.

The linked earlier preprint has a different title and author list from the ICML conference version.

Continue the question

Related reading

Explore our research

Cite this work

Shujun Zhou, Junrong Qi, Guozhang Chen. The Hippocampal Place Field Gradient: A Bio-inspired Framework Building Multiscale Representation for Better Sample Efficiency. ICML 2026.