[Paper Review] Do grid codes afford generalization and flexible decision-making?

Linda Q. Yu, Seongmin A. Park+3 more|arXiv (Cornell University)|Jun 30, 2021

Cognitive Science and Mapping145 references4 citations

TL;DR

This paper investigates how grid codes in the entorhinal cortex may support behavioral flexibility by enabling generalization and adaptive decision-making in abstract, non-spatial tasks. It proposes two complementary mechanisms: grid codes as eigenvectors of the successor representation (SR), providing an efficient basis for learning state transitions, and grid codes as a factorized structural code reflecting global task organization, enabling one-shot inference without new experience.

ABSTRACT

Behavioral flexibility is learning from previous experiences and planning appropriate actions in a changing or novel environment. Successful behavioral adaptation depends on internal models the brain builds to represent the relational structure of an abstract task. Emerging evidence suggests that the well-known roles of the hippocampus and entorhinal cortex (HC-EC) in integrating spatial relationships into cognitive maps can be extended to map the transition structure between states in non-spatial abstract tasks. However, what the EC grid-codes actually compute to afford generalization remains elusive. We introduce two non-exclusive ideas regarding what grid-codes may represent to afford higher-level cognition. One idea is that grid-codes are eigenvectors of the successor representation (SR) learned online during a task. This view assumes that the grid codes serve as an efficient basis function for learning and representing experienced relationships between entities. Subsequently, the grid codes facilitate generalization in novel contexts such as when the goal changes. The second idea is that the grid-codes reflect the inferred global task structure. This view assumes that the grid-code represents a structural code that is factorized from specific sensory content, enabling structural information to be transferred across tasks. Subsequently, the brain could afford one-shot inferences without requiring experience. The ability to generalize experiences and make appropriate decisions in novel situations is critical for both animals and machines. Here we review proposed computations of the grid-code in the brain, which is potentially critical to behavioral flexibility.

Motivation & Objective

• To understand how grid codes in the entorhinal cortex support behavioral flexibility beyond spatial navigation.
• To investigate whether grid codes can represent relational structures in abstract, non-spatial tasks.
• To evaluate whether grid codes serve as an efficient basis for learning state transitions (via successor representation).
• To examine whether grid codes encode a factorized, abstract representation of global task structure for one-shot inference.
• To bridge hippocampal-entorhinal circuit function with higher-level cognition and decision-making in dynamic environments.

Proposed method

• Proposes that grid codes emerge as eigenvectors of the successor representation (SR), a value function that predicts future state occupancy.
• Models grid codes as a low-dimensional, structured basis for representing transition dynamics between states in a task.
• Introduces a computational framework where grid-like representations are learned online during task engagement via temporal difference learning.
• Analyzes how such representations enable generalization when task goals change, by reweighting existing representations.
• Considers grid codes as encoding a structural abstraction of task topology, independent of sensory input or specific rewards.
• Evaluates the potential for one-shot inference by leveraging the factorized, invariant structure of grid codes across different tasks.

Experimental results

Research questions

• RQ1 Can grid codes in the entorhinal cortex serve as an efficient basis for representing state transition dynamics in non-spatial tasks?
• RQ2 To what extent do grid codes reflect the global structural organization of a task, independent of sensory or reward-specific details?
• RQ3 How do grid codes support generalization when task goals or contexts change, without requiring new learning?
• RQ4 Can grid codes enable one-shot inference by abstracting task structure from experience?
• RQ5 What computational mechanisms underlie the emergence of grid-like representations in abstract, non-spatial decision-making?

Key findings

• Grid codes may function as eigenvectors of the successor representation (SR), providing a mathematically efficient basis for learning and generalizing state transition dynamics.
• This SR-based grid code representation enables rapid reconfiguration of behavior when task goals change, supporting flexible decision-making.
• Grid codes can represent a factorized, abstract structure of task topology, allowing transfer of structural knowledge across different tasks.
• The structural abstraction encoded by grid codes may support one-shot inference by enabling the brain to reason about novel task configurations without new experience.
• The dual roles—representing both transition dynamics and global structure—suggest grid codes are central to cognitive flexibility and generalization in abstract reasoning.
• These mechanisms provide a neurocomputational explanation for how the hippocampo-entorhinal system supports adaptive behavior in changing or novel environments.