🧠 neuroscience

Predictive Grid Cell Coding Does Not Require Explicit Planning

This study demonstrates that predictive grid cells, which encode future rather than current locations, are a general feature of the rodent medial entorhinal cortex and emerge naturally within grid cell circuits during path integration, suggesting they may shape population activity rather than solely serving explicit planning functions.

Original authors: Sinha, A., Fasehun, O. C., Redman, W. T.

Published 2026-10-15

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Sinha, A., Fasehun, O. C., Redman, W. T.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Deep within the brain's navigation system lies a small region called the medial entorhinal cortex, a place where the brain builds a mental map of the world. For decades, scientists have known this area contains "grid cells," neurons that fire in a repeating, hexagonal pattern as an animal moves through space, effectively acting as an internal coordinate system. More recently, researchers discovered a special subset of these cells that seem to fire not just where the animal is right now, but where it is about to be. This "predictive" firing suggested the brain might be constantly running ahead, simulating future paths to help with planning a route to a goal. However, a lingering question remained: is this future-looking ability a dedicated tool for planning, or is it simply a byproduct of how the brain's map updates itself?

A team of researchers set out to answer this by looking at the brain in two very different ways: first, by re-examining real recordings from the brains of rats exploring a room, and second, by training artificial neural networks to perform the same task of tracking movement without any instruction to plan ahead. They found that these "predictive grid cells" appear naturally in the real rat brain, even when the animal is simply wandering without a specific destination in mind. More surprisingly, they discovered that the same type of future-looking cells emerged in the artificial networks, even though the networks were never taught to plan or predict the future. The artificial networks were only given the task of updating their position based on movement, yet they spontaneously developed units that fired as if they were looking forward in time.

The researchers tested whether these cells were truly predicting the future or just reacting to the past. They found that the activity of these specific cells depended heavily on the direction the animal was moving. If two paths crossed at the same spot but the animal was heading in different directions, these cells fired differently, showing they were tracking the path ahead, not just the current location. This suggests the cells are encoding a trajectory. The study also revealed that these predictive cells often overlap with another type of cell in the network that looks like a stripe or a band, which helps keep the map stable. This led the authors to propose that these future-looking cells might not be a separate planning module, but rather a fundamental part of the mechanism that keeps the brain's map accurate and moving forward.

To see if this mechanism exists even when the brain is not actively navigating, the team looked at data from rats while they were asleep. They found a faint but clear signature of this predictive coding during sleep, suggesting the brain continues to run these forward-looking simulations even when the animal is resting. The results challenge the idea that the brain needs a specific goal or a conscious plan to generate these future-looking signals. Instead, it appears that the ability to anticipate the next step is built into the very circuitry that allows the brain to know where it is right now. The study suggests that the brain's map is not a static picture of the present, but a dynamic system that is constantly updating itself by projecting a moment into the future, a process that happens whether the animal is planning a journey or simply sleeping.

Technical Summary: Predictive Grid Cell Coding Does Not Require Explicit Planning

Problem Statement
Medial entorhinal cortex (MEC) contains grid cells, which encode spatial location via a hexagonal lattice, and are hypothesized to support path-integration. Recently, a distinct subpopulation termed "predictive grid cells" was discovered; these cells exhibit grid-like tuning not for the animal's current position, but for its future position (prospective coding). While this aligns with hypotheses that MEC supports goal-directed navigation and planning, the functional significance of these cells remains unclear. Existing predictive coding theories often attribute prospective properties to standard grid cells, raising the question of what specific role predictive grid cells play. Furthermore, it is unknown whether this coding is a general feature of rodent MEC or a specific artifact of particular experimental conditions. Crucially, it is unclear if explicit planning mechanisms are required for such representations to emerge, or if they arise naturally from the computational demands of path-integration.

Methodology
The authors employed a dual approach utilizing independent electrophysiological data and recurrent neural network (RNN) modeling:

Electrophysiological Re-analysis: The authors re-analyzed a large-scale, previously published MEC dataset (Gardner et al., 2022) collected from rats in an open-field environment. This dataset was distinct from the original discovery of predictive grid cells (Ouchi & Fujisawa, 2024) in terms of facility, experimental conditions, and analysis code.
- Classification: For 1,330 recorded cells, grid scores were computed at various temporal shifts ( $\Delta$ ). Standard grid cells were defined as having a grid score $>0.5$ at $\Delta=0$ . Predictive grid cells were defined as non-standard grid cells (grid score $<0.5$ at $\Delta=0$ ) that achieved a maximal grid score $>0.3$ at a prospective shift ( $\Delta_{max} > 0$ ).
- Validation: A circular-shift shuffle analysis was performed to ensure that observed grid scores were statistically significant and not due to chance alignment.
- Sleep Analysis: The authors examined whether signatures of predictive grid coding (specifically, bias in toroidal coordinate encoding) persisted during slow-wave and REM sleep.
Recurrent Neural Network (RNN) Modeling: The authors trained 10 independently initialized RNNs to perform path-integration in a 2D environment.
- Architecture: Networks utilized 512 place cells and 4,096 recurrent units, updated via velocity inputs without any explicit goal, route planning, or future-position signals.
- Analysis: The spatial tuning of recurrent units was analyzed using the same temporal shifting methodology applied to the MEC data. Units were classified as standard grid, predictive grid, or retrospective grid based on their grid score profiles across temporal shifts.
- Path-Dependence: To verify prospective coding, the authors computed population vector correlations for pairs of trajectories intersecting the same spatial bin but with different movement vectors.
- Functional Overlap: The authors investigated the overlap between predictive grid units and "band" units (units with band-like spatial tuning known to support path-integration in RNNs).
- Ablation: A preliminary ablation study removed predictive grid units to observe effects on the autocorrelation of the remaining grid unit population.

Key Results

General Presence in MEC: In the independent MEC dataset, the authors identified 117 predictive grid cells (approx. 22.4% of non-standard grid cells per recording). These cells showed low grid scores at $\Delta=0$ but exhibited significant grid-like structure when activity was projected to future positions (temporal shifts of $\approx 100-200$ ms). This confirms predictive grid cells are a general feature of rodent MEC, not limited to the original discovery dataset.
Emergence in Path-Integrating RNNs: Despite being trained only on path-integration with no planning objective, the RNNs spontaneously developed predictive grid units (approx. 4.7% of non-dead units) and retrospective grid units (approx. 3.4%).
- These units showed a monotonic increase in grid score with prospective shifts, distinct from standard grid units which peaked at $\Delta=0$ .
- Predictive grid units in RNNs exhibited significantly higher future path-dependence than standard grid units; their population vectors were highly correlated only when trajectories shared similar future directions, even if they occupied the same current location.
Role of Band Units: The authors found a significant overlap between band units and predictive grid units in RNNs. Band units showed a greater relative increase in grid score with prospective shifts compared to standard grid units, suggesting predictive coding is linked to the circuit mechanisms that update spatial representations.
Sleep Signatures: In one recording where sleep data was available, predictive grid cells (identified during exploration) exhibited a bias toward encoding toroidal coordinates with prospective shifts during both slow-wave and REM sleep, whereas non-predictive cells did not.
Ablation Effects: Removing predictive grid units in RNNs resulted in subtle changes to the grid unit population dynamics, specifically increasing autocorrelation at longer time-scales, suggesting these units may help "repel" activity away from current states in the attractor manifold.

Significance and Claims
The paper argues that predictive grid cell coding is a fundamental property of the MEC grid cell circuit that does not require explicit planning or goal-directed objectives to emerge. The authors challenge the notion that predictive grid cells are solely dedicated to high-level route planning. Instead, they hypothesize that predictive grid cells play a role in maintaining and updating the population dynamics of the grid cell network (e.g., within continuous attractor networks), potentially by driving activity forward along the attractor manifold.

The study provides evidence that:

Predictive grid cells are a general feature of rodent MEC.
Such representations emerge naturally in path-integrating systems without explicit planning objectives.
The presence of predictive coding in band-like units suggests a mechanistic link between path-integration updates and prospective tuning.
Signatures of this coding persist during sleep, supporting the idea that it is an intrinsic property of the circuit rather than a task-specific modulation.

The authors conclude that while downstream regions may leverage these signals for planning, the generation of predictive grid coding itself appears to be a byproduct of the computational requirements for robust path-integration and attractor dynamics.

More like this