Flexible neural representations of abstract structural knowledge in the human Entorhinal Cortex

Shouxiang Ma, Philipp Schwartenbeck, Avital Hahamy, Veronika Samborska, A. Baram, Timothy E.j. BehrensbioRxiv 2023 · DOI

TL;DR

This paper introduces a novel fMRI analysis method, "subspace generalization," to investigate how the human entorhinal cortex (EC) represents abstract structural knowledge. They demonstrate that the EC can generalize representations across different non-spatial tasks that share the same underlying statistical rules (a hexagonal grid structure), even when task size and sensory details differ.

What Problem Does This Paper Address?

Humans excel at generalizing knowledge to new situations, even when those situations differ significantly from prior experiences. This ability is crucial for flexible behavior. While previous research has shown that the entorhinal cortex (EC) plays a role in spatial navigation and abstract task representation, it's unclear how it supports generalization across abstract tasks that share underlying structural rules but vary in other aspects like size or sensory input. Existing fMRI analysis methods often struggle with tasks where there isn't a direct one-to-one mapping between states across different tasks, making it difficult to assess this type of generalization. This paper aims to bridge this gap by developing and validating a new method to probe these abstract, generalizable representations in the EC.

Methods

The core innovation is the "subspace generalization" method. This technique quantifies the similarity between neural representations of different tasks by examining the covariance of neural activity within each task. Specifically:

  1. Covariance Matrix: For each task, a neuron-by-neuron covariance matrix is computed, reflecting how neural activity patterns co-vary across different states within that task.
  2. Principal Components (PCs): These covariance matrices are then decomposed into their principal components (PCs), which represent the dominant patterns of neural activity that explain the most variance in the data. These PCs are ordered by the amount of variance they explain.
  3. Projection and Variance Explained: The neural activity from one task is projected onto the PCs derived from another task. The amount of variance explained by these PCs in the new task is then calculated.
  4. Area Under the Curve (AUC): The cumulative sum of the variance explained across the ordered PCs is plotted, and the area under this curve (AUC) serves as a measure of how well the neural patterns from one task generalize to another. A higher AUC indicates better generalization.

This method is designed to work even when there's no direct mapping between states across tasks, a limitation of methods like Representational Similarity Analysis (RSA). The authors validate this method using electrophysiology data from rodents, simulations, and finally, fMRI data from human participants.

Datasets

  • Rodent Electrophysiology Data: Data from mice freely foraging in two different spatial environments (real and virtual reality) were used to validate the method's ability to capture known grid and place cell generalisation properties.
  • Simulated Data: Simulated grid cells were used to test the method's performance under varying noise levels and voxel resolutions, mimicking fMRI data.
  • Human fMRI Data: Participants performed an abstract associative learning task involving graphs with either hexagonal or community structures. This dataset was used to test the EC's representational generalization capabilities.

Key Insights and Ideas

  • EC Represents Abstract Structure: The EC doesn't just represent spatial layouts; it can encode abstract structural rules, like the hexagonal grid pattern, in a generalizable way. This suggests a more abstract role for the EC than previously thought.
  • Structural Form is Key: The statistical rules governing relationships between elements (the "structural form") are a critical factor for generalization, even more so than the specific sensory details or the size of the environment.
  • Methodological Innovation: The "subspace generalization" method is a powerful new tool for fMRI analysis, enabling the study of neural representations in complex tasks where direct state-to-state mapping is impossible. This opens doors for investigating generalization in a wider range of cognitive domains.
  • Hexagonal vs. Community Structure: Generalization was observed for hexagonal graph structures but not for community structures. This difference highlights the importance of the type of structure and potentially the limitations of the subspace generalization method for certain complex, nested structures.

Key Results

  • Rodent Data Validation: The subspace generalization method successfully captured known differences in generalization between grid cells (good generalization) and place cells (poor generalization, i.e., remapping) in rodents.
  • fMRI Generalization in EC: Human fMRI data showed significant subspace generalization in the EC for hexagonal graphs that differed in size and sensory stimuli. This indicates that the EC represents the underlying hexagonal structure flexibly.
  • Null Result for Community Structure: No clear evidence for subspace generalization was found in the EC for non-spatial tasks with clustered (community) structures, despite behavioral evidence that participants used this structure.

Limitations and Future Work

  • Community Structure Null Result: The lack of generalization for community structures is a key limitation. The authors suggest potential reasons, including the specific nested ring structure used, the nature of the "within-community" vs. "connecting node" representation, and the possibility that subspace generalization might not be sensitive enough for detecting this specific type of structure. Future work could explore different community structures or alternative analysis methods.
  • Generalizability to Other Structures: While the study demonstrates generalization for hexagonal structures, further research is needed to see if this principle extends to other non-Euclidean structural forms (e.g., hierarchical, periodic).
  • EC vs. Other Brain Areas: The paper focuses on EC, but acknowledges evidence for structural representations in other areas like the mPFC. Future work could explore the interplay and dissociation of these representations across different brain regions.
  • Methodological Refinements: The authors suggest that more work is needed to verify the utility of subspace generalization across different neural recording modalities and to understand its sensitivity to noise and signal characteristics.

Summary updated

Abstract

Abstract Humans’ ability for generalisation is outstanding. It is flexible enough to identify cases where knowledge from prior tasks is relevant, even when many features of the current task are different, such as the sensory stimuli or the size of the task state space. We have previously shown that in abstract tasks, humans can generalise knowledge in cases where the only cross-task shared feature is the statistical rules that govern the task’s state-state relationships. Here, we hypothesized that this capacity is associated with generalisable representations in the entorhinal cortex (EC). This hypothesis was based on the EC’s generalisable representations in spatial tasks and recent discoveries about its role in the representation of abstract tasks. We first develop an analysis method capable of testing for such representations in fMRI data, explain why other common methods would have failed for our task, and validate our method through a combination of electrophysiological data analysis, simulations and fMRI sanity checks. We then show with fMRI that EC representations generalise across complex non-spatial tasks that share a hexagonal grid structural form but differ in their size and sensory stimuli, i.e. their only shared feature is the rules governing their statistical structure. There was no clear evidence for such generalisation in EC for non-spatial tasks with clustered, as opposed to planar, structure.
BibTeX
@misc{ma2023flexible,
  author = {Shouxiang Ma and Philipp Schwartenbeck and Avital Hahamy and Veronika Samborska and A. Baram and Timothy E.J. Behrens},
  title = {{Flexible neural representations of abstract structural knowledge in the human Entorhinal Cortex}},
  year = {2023},
  doi = {10.1101/2023.08.31.555760},
  howpublished = {bioRxiv},
}
Figure