Gaussian Process Inference Reveals Non-separability of Positionand Velocity Tuning in Grid Cells

By applying Gaussian Process methods to analyze four-dimensional tuning curves of grid cells in freely foraging rats, this study reveals that position and velocity tuning are often non-separable, a complex interaction that remains undetected in traditional two-dimensional analyses.

The Gist

Imagine your brain has a built-in GPS, a sophisticated internal map that helps you find your way around. For decades, scientists have been studying the "grid cells" in the brain that act like the grid lines on this map. These cells fire in a beautiful, repeating hexagonal pattern as you move, helping you know exactly where you are.

But here's the big question: Does this map stay the same no matter how you move?

If you walk slowly, does the map look the same as when you run? If you move left, is the map identical to when you move right?

For a long time, scientists assumed the answer was "yes." They thought the brain handled where you are (position) and how fast/direction you are moving (velocity) as two completely separate things. It was like thinking the map is a static piece of paper, and your speed just turns a dial that makes the map brighter or dimmer, but never changes the shape of the streets.

This new paper says: Actually, the map is much more dynamic than we thought.

The Problem: A Crowded Room with Empty Chairs

To figure this out, the researchers needed to look at the brain while rats ran around an open field. They wanted to see the "tuning curve"—a fancy term for a map showing how much a brain cell fires for every possible combination of:

1. Where the rat is (Left/Right, Up/Down).
2. How fast and in what direction the rat is moving (Fast Left, Slow Right, etc.).

This creates a 4-dimensional space (Position X, Position Y, Velocity X, Velocity Y).

The Analogy: Imagine trying to take a photo of a crowded dance floor, but you only get to take a picture of the dancers every time they land on a specific tile. Because rats don't run in straight lines at constant speeds, they skip over huge chunks of the dance floor. Some tiles are visited thousands of times; others are never touched.

If you tried to draw a map based only on the tiles the rats actually visited, you'd have a map full of giant holes. You couldn't tell if the map was changing or if you just hadn't looked hard enough.

The Solution: The "Magic Predictor" (Gaussian Processes)

To fill in the holes, the researchers used a statistical tool called a Gaussian Process (GP).

The Analogy: Think of the GP as a super-smart, artistic predictor. If you show it a few dots on a piece of paper, it doesn't just connect the dots with straight lines; it uses the pattern of the dots to guess what the picture looks like in the empty spaces. It says, "Based on the dots here and there, I'm 90% sure there's a curve here, and maybe a hill over there."

By using this "Magic Predictor," the researchers could fill in the gaps of the rats' movement data and create a smooth, complete 4D map of how the brain cells were firing.

The Discovery: The Map Changes Shape

Once they had the complete maps, they compared them to the "old school" theory (that position and speed are separate).

• The Old Theory (Separable): Imagine a projector showing a map on a wall. If you turn up the speed, the projector just makes the image brighter. The streets don't move; they just glow more.
• The New Reality (Non-Separable): The researchers found that for many brain cells, the map actually changes shape. When the rat runs fast, the "streets" on the internal map might shift, stretch, or even disappear and reappear in new places.

The Metaphor: It's like driving a car.

• Separable view: The road signs stay in the same place whether you drive at 20 mph or 60 mph; you just see them faster.
• Non-separable view: At 20 mph, the road looks like a straight highway. At 60 mph, the road suddenly curves, or a new exit appears. The road itself is reacting to your speed.

Why Does This Matter?

The researchers found that this "shape-shifting" only became obvious when they had enough data to fill in the map. In sessions where the rats ran around a lot (covering more of the dance floor), the "Magic Predictor" (GP) showed that the brain cells were doing something complex and interactive. In sessions where the rats didn't move around enough, the data was too sparse, and it looked like the old, simple theory was correct.

The Takeaway:
Our brain's GPS isn't just a static map with a volume knob for speed. It's a living, breathing, 4D representation where where you are and how you are moving are deeply intertwined. The map flexes and shifts based on your movement, suggesting our brains are constantly recalculating our reality in real-time, not just tracking a fixed location.

This study is a reminder that to understand the complex machinery of the brain, we sometimes need better tools to fill in the blanks, because the truth is often hidden in the spaces we didn't think to look.

Technical Summary

1. Problem Statement

Grid cells in the medial entorhinal cortex (MEC) are fundamental to spatial navigation, traditionally characterized by their periodic, hexagonal firing patterns in 2D space. While it is established that grid cells also encode variables like speed and head direction, the nature of the interaction between position and velocity remains unclear.

• The Core Question: Do grid cells encode position and velocity independently (a separable code where velocity acts merely as a gain modulator on a static position map), or do they integrate these variables into a complex, non-separable 4D representation where the position tuning itself changes based on velocity?
• The Challenge: Investigating this requires comprehensive sampling of a 4D behavioral space (2D position $\times$ 2D velocity). In natural foraging behavior, animals rarely visit all combinations of position and velocity, leading to extreme data sparsity in the 4D space. Traditional binning methods fail to estimate firing rates at unvisited points, making it difficult to distinguish true non-separability from sampling artifacts.

2. Methodology

The authors leveraged a large-scale dataset from Gardner et al. (2022) containing Neuropixels recordings from freely foraging rats in an open field.

• Data Preprocessing:

  • Dimensions: Constructed a 4D behavioral space comprising $x$-position, $y$-position, $x$-velocity, and $y$-velocity.
  • Binning: The space was discretized into $30 \times 30$ position bins (5 cm) and $5 \times 5$ velocity bins (10 cm/s), resulting in 22,500 total bins.
  • Sparsity Issue: Analysis revealed that in most sessions, rats visited fewer than 50% of these bins, creating significant gaps in the data.

• Gaussian Process (GP) Inference:

  • To overcome sparsity, the authors applied Gaussian Process (GP) regression to estimate firing rates at unobserved points in the 4D space.
  • Kernel: Used a Matérn kernel ($\nu = 5/2$) with learned length scales (initialized at 10 for position, 10 cm/s for velocity).
  • Validation: Employed $k$-fold cross-validation (80/20 split) to ensure the model was not hallucinating patterns. The model's performance was measured using the Fraction of Variance Explained (FVE) on held-out data.
  • Null Model: Simulated spike trains using a null model (invariant tuning based on zero-velocity data) to confirm that observed non-separability was not an artifact of the GP interpolation.

• Comparative Modeling:

  • Separable Model: A baseline model assuming the firing rate $r(x, y, v_x, v_y)$ is the product of independent position and velocity functions: $f_{pos}(x, y) \cdot f_{vel}(v_x, v_y)$. This was optimized using gradient descent (PyTorch) with non-negativity constraints.
  • Comparison: The authors compared the FVE of the full GP model against the separable model on held-out data. A significant improvement in FVE by the GP model indicates non-separability.

• Quantitative Metrics:

  • $\Delta$FVE: Defined as $FVE_{GP} - FVE_{separable}$. Positive values suggest non-separability.
  • SDCS (Standard Deviation of Cosine Similarity): Calculated the cosine similarity between position tuning curves at specific velocity bins and the marginalized position tuning curve. High SDCS indicates that the shape of the grid fields changes significantly across velocities.

3. Key Results

• Evidence of Non-Separability:

  • In sessions with high data density (specifically Rat R, Day 1), the GP model significantly outperformed the separable model ($\Delta$FVE > 0). This indicates that for many cells, position tuning is not static but interacts with velocity.
  • Visualizations of single-cell tuning curves revealed that grid fields can shift location, change shape, appear, or disappear depending on the velocity vector, rather than just scaling in magnitude (gain modulation).

• Data Coverage Threshold:

  • The ability to detect non-separability is strictly dependent on data coverage. In sessions with sparse data (shorter durations), the separable model often performed as well as or better than the GP model.
  • A strong positive correlation was found between data density (fraction of bins visited) and $\Delta$FVE. This suggests that previous studies failing to find non-separability may have been limited by insufficient behavioral sampling.

• Cellular Variability:

  • Non-separability is not uniform across all cells. Some cells showed gradual changes in tuning, while others exhibited drastic shifts (e.g., grid field disappearance).
  • Correlations: $\Delta$FVE was significantly correlated with mean firing rate and SDCS, but not with grid score or grid scale. This implies that the phenomenon is driven by the richness of the neural signal and the complexity of the tuning interaction, not by the geometric quality of the grid itself.

• Validation:

  • Simulations using the null model (where tuning was truly separable) showed that the GP model did not generate false non-separability when data was sufficient, confirming the robustness of the findings.

4. Key Contributions

1. Methodological Innovation: Successfully applied Gaussian Processes to reconstruct 4D neural tuning curves (Position $\times$ Velocity) from sparse behavioral data, overcoming a major limitation in high-dimensional neuroscience analysis.
2. Theoretical Insight: Provided strong evidence that grid cell coding is non-separable. Position and velocity are integrated into a joint 4D neural manifold, challenging the view that velocity acts solely as a multiplicative gain on a static spatial map.
3. Identification of Sampling Bias: Established a critical data coverage threshold required to observe non-separable tuning. This explains discrepancies in the literature and sets a standard for future high-dimensional neural recording studies.
4. Characterization of Tuning Dynamics: Demonstrated that velocity dependence manifests as complex structural changes in grid fields (translation, distortion, appearance/disappearance) rather than simple gain modulation.

5. Significance

• Revising Neural Navigation Models: The findings suggest that computational models of the MEC, such as Continuous Attractor Networks (CANs), must account for dynamic, velocity-dependent shifts in the spatial representation, rather than assuming a rigid metric.
• Mechanistic Implications: The non-separability implies that the neural circuitry underlying grid cells may involve mechanisms like traveling waves, prospective coding, or theta sweeps that integrate speed and direction directly into the spatial code.
• Future Directions: This framework provides a blueprint for analyzing even higher-dimensional representations (e.g., adding pupil size, whisking, or task context) to uncover interaction factors that are averaged out in lower-dimensional analyses. It underscores the necessity of high-dimensional modeling to fully understand the neural code for external space.