[Paper Review] Generalisation of structural knowledge in the hippocampal-entorhinal system

James C. R. Whittington, Timothy Müller+3 more|arXiv (Cornell University)|May 23, 2018

Memory and Neural Mechanisms24 references22 citations

TL;DR

This paper proposes that generalisation of structural knowledge in the brain arises from separating spatial structure from sensory stimuli, using hierarchical grid-like representations in the entorhinal cortex and conjunctive place-cell codes that bind structure to sensory input. In an artificial neural network, this mechanism enables zero-shot inference and hierarchical memory addressing, with experimental data confirming that grid-place cell relationships are preserved across environments—supporting a unified basis for structural generalisation.

ABSTRACT

A central problem to understanding intelligence is the concept of generalisation. This allows previously learnt structure to be exploited to solve tasks in novel situations differing in their particularities. We take inspiration from neuroscience, specifically the hippocampal-entorhinal system known to be important for generalisation. We propose that to generalise structural knowledge, the representations of the structure of the world, i.e. how entities in the world relate to each other, need to be separated from representations of the entities themselves. We show, under these principles, artificial neural networks embedded with hierarchy and fast Hebbian memory, can learn the statistics of memories and generalise structural knowledge. Spatial neuronal representations mirroring those found in the brain emerge, suggesting spatial cognition is an instance of more general organising principles. We further unify many entorhinal cell types as basis functions for constructing transition graphs, and show these representations effectively utilise memories. We experimentally support model assumptions, showing a preserved relationship between entorhinal grid and hippocampal place cells across environments.

Motivation & Objective

• To understand how the hippocampal-entorhinal system enables generalisation of structural knowledge across novel situations.
• To address the limitation in artificial intelligence where statistical patterns are learned without explicit structural abstraction.
• To unify diverse entorhinal cell types (grid, border, object vector) as basis functions for transition statistics.
• To test whether place cell remapping is non-random and consistent with structural knowledge, rather than random or stimulus-driven.
• To demonstrate that hierarchical, unsupervised learning in artificial neural networks can generate brain-like representations that support zero-shot inference and memory addressing.

Proposed method

• The model uses an artificial neural network trained in a purely unsupervised manner to predict sensory observations on 2D graph worlds with structured connectivity.
• Grid-like representations emerge in the entorhinal layer as basis functions for transition statistics, encoding the underlying structural geometry of the graph.
• Place cells form a conjunctive Hebbian memory by combining grid-based structural codes with sensory inputs, enabling fast, context-specific associations.
• A hierarchical architecture enables efficient combinatorial coding, allowing the system to generalise to unseen graph structures.
• The model uses fast Hebbian learning to maintain short-term associations between sensory stimuli and structural representations.
• Experimental validation involves computing gridAtPlace and minDist metrics to quantify grid-place cell relationships across environments, comparing with null distributions from shuffled data.

Experimental results

Research questions

• RQ1 Can structural knowledge be generalised by separating it from sensory stimuli in a neural system?
• RQ2 Do grid cells preserve their relationship with place cells across different environments, indicating structural consistency?
• RQ3 Is place cell remapping non-random and consistent with underlying structural knowledge rather than purely sensory-driven?
• RQ4 Can artificial neural networks with hierarchical, Hebbian memory learn to generalise structural statistics without supervision?
• RQ5 Do unified basis functions (e.g. grid, border, object vector cells) collectively represent transition statistics in a way that supports generalisation?

Key findings

• The gridAtPlace measure showed a significant correlation (p < 0.005) between place and grid cells across different environments, indicating preserved functional relationships.
• The minimum distance (minDist) measure also showed significant correlation across environments, supporting non-random remapping.
• The correlation of grid cell activity patterns across environments was preserved (p < 0.001), indicating stable structural representation.
• Model-generated rate maps reproduced the observed correlation of 0.3–0.35 between grid and place cells, matching empirical data.
• The grid cell correlation matrix was significantly correlated across environments in both data and model, confirming structural invariance.
• The results support that place cells remap in a non-random, structure-consistent manner, validating the model’s core assumption of conjunctive coding.