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

The hippocampus encodes space through a striking gradient of place field sizes along its dorsal-ventral axis, yet the principles generating this continuous gradient from discrete grid cell inputs remain unclear. We propose a unified theoretical framework establishing how multiscale hippocampal place fields arise from the frequency-dependent decay of grid cell projections. Functionally, this organization establishes an inductive bias in the population code, managing a fundamental trade-off between spatial precision and sample efficiency. Translating this insight to artificial neural networks, we incorporate a hippocampus-inspired positional embedding (HIPE) into the Transformer architecture to induce multi-scale representation. Experimental results confirm that this mechanism effectively improves data efficiency. Our work establishes a link between neural connectivity, activity patterns, and learning, suggesting a principled approach to utilizing multi-scale representations for sample-efficiency learning. Our codes are available at https://github.com/AIogry/relative_PE.

The hippocampus, a brain region important for memory and navigation, contains “place cells” that become active in particular locations. Some place cells respond to small, precise areas, while others respond across much broader areas. However, how this range of scales emerges, and why it is useful, has remained unclear. We use mathematics to show how place cells with different firing-field sizes can arise from structured connections with grid cells, another type of spatially responsive neuron. We further show that this multiscale design supports learning from very few examples: broad representations help generalize from limited experience, while precise representations preserve fine detail. We then transfer this shared principle to artificial intelligence by developing HIPE, a brain-inspired positional representation for Transformers. In experiments, HIPE helped models learn more effectively when training data were limited. Our work suggests that representing information across multiple scales may be a common strategy for sample-efficient learning in both biological and artificial intelligence.