Abstract
Grid cells have been proposed to encode task regularities that allow predicting future states. Entorhinal grid-like signals might therefore mirror behavioral biases associated with relying on task regularities, like regression-to-the-mean biases in time estimation. Here, we tested this proposal using functional magnetic resonance imaging and a rapid timing task in humans. Indeed, trial-wise entorhinal activity reflected task accuracy and the degree to which interval estimates regressed towards the mean of all tested intervals. Grid-like signals were observed exclusively for the interval closest to the mean, which was explained by differences in temporal stability across intervals. Finally, both behavioral and entorhinal results were explained by a Bayesian observer model that assumes the integration of current-trial sensory evidence with prior expectations. Together, we find that entorhinal activity and grid-like signals reflect behavioral performance in a timing task, supporting the proposed role of grid cells in encoding task structure for predictive coordination of behavior.
Introduction
Grid cells in the entorhinal cortex are well-known for providing a coordinate system for spatial navigation. However, recent theories suggest a broader role: encoding the "structure" of any task, whether spatial or non-spatial. A prominent feature of human cognition is the regression-to-the-mean bias, where estimates of magnitudes (like duration) are pulled toward the average of the sampled distribution. This bias is considered an optimal Bayesian strategy for managing sensory uncertainty. This study investigates whether entorhinal grid-like signals reflect these behavioral timing biases, potentially serving as a neural substrate for encoding task regularities.
Methods
The researchers employed high-resolution fMRI (3T Siemens Magnetom Skyra) while 34 healthy participants performed a Time-to-Contact (TTC) task.
Experimental Design
Participants tracked a dot moving at one of four velocities toward a boundary. After the dot disappeared, participants estimated when it would hit the boundary, resulting in four target intervals: 0.55s, 0.67s, 0.86s, and 1.2s. Feedback was provided after each trial.
Analysis and Modeling
Neural analysis focused on "hexadirectional modulation" in the posteromedial entorhinal cortex (pmEC) as a proxy for grid cell activity. A Bayesian observer model was used to simulate the integration of current-trial sensory evidence ("Likelihood") with the distribution of previously encountered intervals ("Prior").
Results
Behavioral Biases
Participants exhibited a clear regression-to-the-mean effect. Estimates for short intervals were over-estimated, while long intervals were under-estimated, resulting in a regression slope of 0.53. Root-mean-square error (RMSE) followed a quadratic pattern, being lowest for intervals near the mean of the distribution (0.82s).
Neural Correlates of Task Regularity
Activity in the pmEC significantly correlated with both trial-wise accuracy ($p = 1.8 \times 10^{-4}$) and the magnitude of the regression-to-the-mean bias ($p = 0.015$). Crucially, reliable 6-fold symmetric grid-like signals were observed exclusively for the interval closest to the mean (0.86s; $p = 0.029$). This specificity was attributed to higher temporal stability of neural representations when task demands aligned with internal prior expectations.
Computational Validation
The Bayesian observer model successfully replicated both the behavioral results (Mean Absolute Error = 0.02) and the neural patterns. It demonstrated that the strongest and most stable neural signals occur when sensory evidence aligns with the learned statistical structure of the task.
Discussion
These findings support the hypothesis that grid cells encode task regularities to guide behavior. The link between pmEC activity and Bayesian biases suggests that the entorhinal cortex builds a "map" of the statistical environment.
Limitations and Future Directions
The study's reliance on four discrete intervals limits the ability to see how grid signals adapt to shifting distributions. Future research should explore a wider range of intervals and test the causal role of these signals using non-invasive stimulation or by studying patients with entorhinal damage. Additionally, investigating whether this mechanism generalizes to other non-spatial domains (e.g., loudness or social hierarchy) remains a key area for exploration.
Summary updated