Structure Learning in Bayesian Sensorimotor Integration

Tim Genewein, Eduard Hez, Zeynab Razzaghpanah, Daniel A. Braun

Introduction

In sensorimotor control, the human brain must constantly integrate internal expectations (priors) with sensory feedback to navigate complex environments. Traditional Bayesian models of motor learning typically focus on how we estimate independent variables, such as the position of a single point. However, real-world environments often contain "hidden" structural relationships where variables are statistically linked. For instance, a gust of wind affecting a tennis ball's lateral trajectory likely also influences its depth. This study investigates whether the human motor system can learn these underlying correlations—a process termed Bayesian structure learning—to infer missing information in one dimension based on feedback from another.

Methodology

To test this hypothesis, researchers employed a 3D virtual reality environment where participants used a robotic manipulandum to reach for targets. A hidden 2D translational shift was applied to the cursor representing the hand's position. Participants were divided into two primary groups:

  • Correlated Group: Experienced shifts where horizontal and vertical displacements were highly correlated (e.g., a leftward jump predicted an upward jump).
  • Uncorrelated Group: Experienced random, independent shifts in both dimensions.

A critical component of the study was the "Partial Feedback" condition. In specific trials, participants received feedback only in the horizontal dimension via a flickering vertical bar, providing no direct information about the vertical error. To succeed, participants had to rely on their learned internal model of the environment's structure. The researchers modeled this behavior using a hierarchical Bayesian framework that learns both the mean and the covariance (the structure) of errors over time using a Normal Inverse-Wishart distribution.

Key Insights

The study yielded several significant findings regarding the nature of motor learning:

  • Temporal Dynamics of Learning: While participants learned the mean shift almost immediately, learning the correlation (the structure) was a significantly slower "meta-learning" process. It required thousands of trials over four days to manifest.
  • Implicit vs. Explicit Learning: In a control experiment, a group explicitly informed of the correlation performed worse than those who learned it implicitly. This suggests that explicit cognitive strategies may interfere with the natural, implicit processes of the motor system.
  • Bounded Rationality and Sampling: In trials with zero feedback, participants' errors were not merely random but reflected the learned correlation. This behavior suggests the brain does not simply perform Maximum A Posteriori (MAP) estimation but instead "samples" from its internal belief distribution, a characteristic of bounded rationality.
  • The Necessity of Performance Feedback: Structure learning required binary "hit/miss" feedback. A group that observed the shifts but never received performance outcomes failed to learn the underlying correlation.

Results

By the fourth day of training, the correlated group demonstrated a significant reduction in error during partial feedback trials. Their "slope"—a metric indicating the degree of reliance on the prior versus feedback—dropped from 1.0 (no learning) to approximately 0.5–0.6. In contrast, the uncorrelated group showed no such improvement, confirming that the motor system specifically adapts to the statistical structure of the environment. The hierarchical Bayesian model closely matched these behavioral results, capturing both the rapid acquisition of the mean and the gradual learning of the covariance structure.

Discussion and Limitations

The findings suggest that the human motor system is capable of sophisticated statistical inference, learning not just average conditions but the complex relationships between environmental variables. However, even after 4,000 trials, participants had not reached optimal performance, indicating that internalizing complex environmental statistics is a long-term process.

Future research is needed to determine how these learned structures generalize across different tasks and to identify the specific neural or computational mechanisms (such as Monte Carlo sampling) that the brain uses to implement these Bayesian inferences.

Summary updated

Abstract

Author Summary The human sensorimotor system has to process highly structured information that is affected by uncertainty and variability at all levels. Previously, it has been shown that sensorimotor processing is very efficient at extracting structure even in variable environments and it has also been shown how sensorimotor integration takes into account uncertainty when processing novel information. In particular, the latter integration process has been shown to be consistent with Bayesian theory. Here we show how the two processes of structure learning and sensorimotor integration work together in a single experiment. We find that when human participants learn a novel motor skill they not only successfully extract structural knowledge from variable data, but they also exploit this structural knowledge for near-optimal sensorimotor integration.
BibTeX
@article{genewein2015structure,
  author = {Tim Genewein and Eduard Hez and Zeynab Razzaghpanah and Daniel A. Braun},
  title = {{Structure Learning in Bayesian Sensorimotor Integration}},
  journal = {PLOS Computational Biology},
  year = {2015},
  doi = {10.1371/journal.pcbi.1004369},
}
Figure