Encoding of task regularities links grid-like signals to human timing behavior

This study demonstrates that entorhinal grid-like signals in humans encode task regularities by reflecting both behavioral regression-to-the-mean biases in time estimation and the integration of sensory evidence with prior expectations, thereby supporting the role of grid cells in predictive processing.

The Gist

The Big Idea: How Your Brain Predicts the Future

Imagine you are playing catch in the park. A ball is thrown toward you. To catch it, your brain has to guess exactly when it will arrive.

If you only looked at the ball's current speed and distance, you might miss. But your brain is smarter than that. It remembers that in the last 10 throws, the ball usually arrived in about 1 second. So, even if this specific ball is moving slightly faster or slower, your brain "leans" your guess toward that familiar 1-second mark.

This paper is about where in the brain this "leaning" happens and how the brain does it. The researchers found that a tiny, deep part of your brain called the Entorhinal Cortex acts like a super-organized map, helping you predict time just as it helps you navigate space.

---

The Experiment: The "Invisible Ball" Game

The researchers put 34 people in an MRI machine (a giant camera that takes pictures of the brain) and played a game with them.

1. The Game: A dot moved across the screen. Then, it disappeared behind a wall (occlusion).
2. The Challenge: The participants had to guess exactly when the dot would hit the other side of the wall.
3. The Trick: The dot moved at different speeds, meaning the time it was hidden varied (some were very short, some were long). However, the participants didn't know the rules; they just learned by playing hundreds of rounds.

The Result: The participants were good at the game, but they made a specific mistake. When the hidden time was very short, they guessed it was a bit longer. When it was very long, they guessed it was a bit shorter. They were all "regressing to the mean"—pulling their guesses toward the average time of all the rounds they had played.

The Discovery: The Brain's "Grid"

The researchers looked at the participants' brains while they played. They focused on the Entorhinal Cortex, a region famous for containing Grid Cells.

• What are Grid Cells? Think of them as the brain's GPS. In a city, you have a grid of streets (North-South, East-West). Grid cells create a similar hexagonal (honeycomb) map in your brain to help you know where you are in space.
• The Surprise: Usually, we think these cells only care about space (where you are). But this study found that in this timing game, these cells were also creating a "map" of time.

The "Sweet Spot" Analogy

Here is the most fascinating part of the discovery.

Imagine the different time intervals the participants had to guess were like different musical notes.

• Some notes were very high (fast times).
• Some were very low (slow times).
• One note was right in the middle (the average time).

The researchers found that the brain's "Grid Signal" (the honeycomb map) was strongest and most stable only for that one "middle note" (the average time).

The Analogy: Imagine you are trying to tune a radio.

• When you try to tune into a station that is very far from your favorite station (the average), the signal is fuzzy and jumps around.
• But when you tune into your favorite station (the average), the signal locks in perfectly. The static disappears, and the music is clear.

In the brain, the "Grid Cells" locked in perfectly only when the time interval was close to the average of all the previous rounds. When the time was weird or extreme, the grid signal got shaky and unstable.

Why Does This Happen? (The Bayesian Chef)

The paper suggests the brain works like a Bayesian Chef.

1. The Ingredients (Sensory Evidence): You see the dot moving. That's your current evidence.
2. The Recipe Book (Prior Expectations): You have a mental recipe book built from all your past experiences. It says, "Usually, this dot takes about 1 second."
3. The Dish (The Prediction): To make the best guess, the Chef mixes the current evidence with the recipe book.

If the current evidence is a bit blurry (maybe the dot moved fast), the Chef relies more on the Recipe Book. This causes the "regression to the mean"—your guess gets pulled toward the average.

The study found that the Grid Cells in the brain are the ones doing this mixing. They represent the "Recipe Book." When the current situation matches the recipe (the average time), the Grid Cells fire strongly and steadily. When it doesn't match, they get confused.

The Takeaway

This paper tells us that the part of your brain responsible for knowing where you are (the Entorhinal Cortex) is also the part responsible for knowing when things will happen.

It uses a "grid" system not just to map the streets of a city, but to map the flow of time. And just like a GPS works best when you are on a familiar road, this brain grid works best when the timing is familiar (close to the average). This helps us make quick, smart predictions about the world, even when things are a little uncertain.

In short: Your brain has a built-in honeycomb map that helps you predict the future, and it works best when you are dealing with the "average" of what you've seen before.

Technical Summary

1. Problem and Research Question

The study addresses a fundamental question in cognitive neuroscience: How does the brain encode statistical regularities to support predictive processing and behavior?

• Context: While grid cells in the entorhinal cortex (EC) are well-known for encoding spatial maps, recent theories suggest they may also encode non-spatial task structures (e.g., temporal or abstract relationships) to facilitate generalization and prediction.
• Gap: It remains unclear whether grid-like signals in the human EC are merely epiphenomena of spatial navigation or if they actively reflect the encoding of task regularities (such as temporal intervals) that guide behavior in real-time.
• Specific Hypothesis: The authors hypothesized that entorhinal activity, specifically grid-like signals, would reflect behavioral biases associated with learning task regularities. Specifically, they tested if these signals correlate with the "regression-to-the-mean" bias observed in time estimation tasks, where estimates are biased toward the average of previously experienced intervals.

2. Methodology

Experimental Design

• Task: A rapid Time-to-Contact (TTC) estimation task combined with a visual tracking component.
  • Visual Tracking: Participants tracked a moving fixation target for 10 degrees of visual angle (dva) in one of 24 directions at one of 4 speeds.
  • Occlusion & Prediction: The target was then occluded for 5 dva. Participants had to predict the time ($TTC_e$) the target would hit a boundary.
  • Intervals: Four target TTCs ($TTC_t$) were tested: 0.55s, 0.65s, 0.86s, and 1.2s. The mean of these intervals was 0.82s.
  • Feedback: Participants received feedback on accuracy after each trial.
• Participants: 34 healthy adults (after exclusions) underwent fMRI scanning.
• Data Acquisition: 3T fMRI (Siemens Skyra) with high temporal resolution (TR=1020ms) to capture trial-wise dynamics. Eye-tracking (500 Hz) was used to monitor gaze direction.

Analytical Approaches

1. Behavioral Analysis:
   • Used Linear Mixed-Effects Models (LME) to quantify the relationship between estimated TTC ($TTC_e$) and target TTC ($TTC_t$).
   • Calculated Root-Mean-Square Error (RMSE), decomposed into Variability (precision) and Bias (accuracy relative to ground truth).
2. fMRI Analysis (GLM):
   • Parametric Modulation: Modeled posteromedial entorhinal cortex (pmEC) activity as a function of (a) trial-wise accuracy and (b) the magnitude of the regression-to-the-mean bias.
   • Hexadirectional Grid Analysis: Used a cross-validation approach to detect grid-like signals (6-fold rotational symmetry) in the pmEC.
     • Step 1: Estimated putative grid orientation in half the data (odd runs) based on gaze direction ($\sin(6\theta)$ and $\cos(6\theta)$).
     • Step 2: Tested the amplitude of the grid signal in the held-out data (even runs) by aligning gaze direction to the estimated orientation.
   • Stability Analysis: Assessed spatial stability (clustering of grid orientations across voxels) and temporal stability (consistency of grid orientation across data partitions) for each interval.
3. Computational Modeling:
   • Implemented a Bayesian Observer Model (Bayes Least-Squares, BLS).
   • The model integrated Sensory Likelihood (noisy current trial evidence) with a Prior (Gaussian distribution of learned intervals).
   • Compared BLS against Maximum Likelihood Estimation (MLE) and Prior-dominant (PRI) models using Bayesian Information Criterion (BIC).

3. Key Results

Behavioral Findings

• Regression-to-the-Mean: Participants' estimates were systematically biased toward the mean interval (0.82s). The slope of the regression line was 0.53 (significantly less than 1.0, indicating perfect performance).
• Performance Pattern: RMSE showed a quadratic relationship with the target TTC, being lowest near the mean (0.82s) and higher for extreme intervals. This indicates that participants relied on prior knowledge accumulated across trials.

Neuroimaging Findings (pmEC Activity)

• Trial-Wise Activity: pmEC activity was higher in trials with:
  1. Higher accuracy (lower error).
  2. Stronger regression-to-the-mean bias (larger deviation from the identity line toward the mean).
• Grid-Like Signals:
  • A reliable hexadirectional (6-fold) modulation was observed in the pmEC, but only for the interval closest to the mean ($TTC_{0.86}$).
  • No significant grid-like signals were found for the other three intervals or in control regions (V1, 4-fold/8-fold symmetries).
• Mechanism of Signal Strength: The stronger grid signal for the mean interval was driven by temporal stability (consistency of grid orientation over time), not spatial stability.
• Behavioral Correlation: Within participants, a stronger grid-like signal amplitude predicted lower RMSE (better performance). This correlation was specific to the pmEC and not observed in the visual cortex.

Computational Modeling

• The Bayesian Observer Model (BLS) successfully replicated the behavioral data (quadratic RMSE pattern and regression bias), outperforming MLE and PRI models.
• Crucially, the model-derived RMSE correlated with the amplitude of the observed grid-like signals, suggesting the grid signal reflects the integration of prior expectations and sensory evidence.
• Model fitting indicated that performance improvements were driven by learning the temporal context (sharpening the prior distribution) rather than simply reducing sensory noise.

4. Key Contributions

1. Linking Grid Cells to Non-Spatial Timing: Demonstrates that human entorhinal grid-like signals are not limited to spatial navigation but are modulated by temporal task regularities.
2. Real-Time Encoding of Task Structure: Shows that pmEC activity tracks trial-by-trial behavioral biases (regression-to-the-mean) in real-time, supporting the theory that the EC encodes task structure for predictive processing.
3. Stability as a Mechanism: Identifies temporal stability of the grid orientation as the critical factor determining the strength of the grid-like signal, specifically for the most predictable (mean) interval.
4. Computational Validation: Provides a unified normative explanation using a Bayesian framework, linking behavioral biases, neural grid signals, and the integration of priors with sensory evidence.

5. Significance

• Theoretical Impact: This study supports the "cognitive map" theory, extending the role of the entorhinal cortex from a spatial map to a general predictive coding engine that encodes statistical regularities across domains (space, time, abstract features).
• Mechanistic Insight: It suggests that the brain uses grid-like representations to stabilize predictions about future states (in this case, time-to-contact) based on learned environmental statistics.
• Clinical Relevance: Given the involvement of the EC in Alzheimer's disease and the importance of timing in many cognitive disorders, understanding how these regions encode task regularities could inform future diagnostics and interventions.
• Methodological Advance: The study successfully applies cross-validated grid analysis to a rapid, non-spatial timing task, opening new avenues for investigating domain-general functions of the hippocampal formation.

In conclusion, the paper provides compelling evidence that entorhinal grid-like signals serve as a neural substrate for encoding task regularities, facilitating predictive timing behavior through the integration of prior expectations and current sensory evidence.