A sensorimotor paradigm for Bayesian model selection

Tim Genewein, Daniel A. BraunFrontiers in Human Neuroscience 2012 · DOI

Abstract

Sensorimotor control is thought to rely on predictive internal models in order to cope efficiently with uncertain environments. Recently, it has been shown that humans not only learn different internal models for different tasks, but that they also extract common structure between tasks. This raises the question of how the motor system selects between different structures or models, when each model can be associated with a range of different task-specific parameters. Here we design a sensorimotor task that requires subjects to compensate visuomotor shifts in a three-dimensional virtual reality setup, where one of the dimensions can be mapped to a model variable and the other dimension to the parameter variable. By introducing probe trials that are neutral in the parameter dimension, we can directly test for model selection. We found that model selection procedures based on Bayesian statistics provided a better explanation for subjects' choice behavior than simple non-probabilistic heuristics. Our experimental design lends itself to the general study of model selection in a sensorimotor context as it allows to separately query model and parameter variables from subjects.

Introduction

In sensorimotor control, the brain utilizes "internal models" to predict the consequences of actions and compensate for environmental perturbations, such as the weight of a tool or a visual distortion. While much research has focused on parameter adaptation—the process of adjusting to a specific value within a known rule (e.g., shifting aim by exactly 5cm to compensate for prism glasses)—a more fundamental challenge is model selection.

Model selection involves deciding which underlying structure or rule applies to a given situation. For instance, is a visual shift caused by a stable offset or a random fluctuation? This study investigates whether the human motor system employs probabilistic Bayesian logic to navigate this structural uncertainty, particularly when sensory feedback is ambiguous.

Methods and Experimental Design

The researchers developed a 3D reaching task using a Sensable Phantom Premium haptic manipulandum and a head-mounted display for virtual reality. Seven naive subjects participated in the study, each performing 1,000 trials.

The Task Structure

Subjects moved a cursor toward targets in a 3D environment. During the reach, a lateral shift ($s$) was applied to the cursor. This shift was drawn from one of two distributions:

  • Model 1 ($M_1$): Small shifts centered around zero (Gaussian).
  • Model 2 ($M_2$): Large shifts (a mixture of two Gaussians centered at $\pm 2.5$ cm).

In standard trials, subjects saw a brief flash of the shifted cursor and had to hit the correct target ($M_1$ corresponded to the Upper target, $M_2$ to the Lower target) while compensating for the lateral shift.

The experimental setup: (A) Standard trials with single-point feedback; (B) Probe trials with ambiguous feedback arrays.

Innovation: Probe Trials

To isolate model selection from parameter compensation, the researchers introduced "probe trials." In these trials, the feedback was not a single point but an ambiguous array of rectangles representing a range of possible shifts. Crucially, the targets in these trials were widened so that subjects did not need to compensate for the shift; they only had to choose which target (and thus which model) they believed was most likely given the ambiguous evidence.

Computational Modeling

Five strategies were compared to explain subject behavior:

  1. Bayes Factors: Selection based on the marginal likelihood (how well the model explains the average of all possible shifts in the probe).
  2. Bayesian Policy Inference: A weighted superposition of policies learned during standard trials.
  3. Heuristics: Three non-probabilistic rules (assuming the shift is the "average," the "biggest," or "halfway" across the ambiguous array).

Key Results

Bayesian Superiority

The experimental data strongly favored Bayesian explanations. Both the Bayes Factors and Bayesian Policy Inference models provided significantly better fits to subject behavior than any of the three heuristics. Specifically, the Bayesian Policy Inference model achieved the best fit with a total negative log-likelihood of $L = 1097$, whereas the best heuristic ("Halfway Shift") trailed at $L = 1203$.

Sensitivity to Model Complexity

Subjects demonstrated a sensitivity to the "Occam’s Razor" principle inherent in Bayesian inference. In the second half of the experiment, the researchers increased the variance of $M_1$. Subjects adjusted their choices for the exact same ambiguous stimulus, shifting away from $M_1$ as it became a less "convincing" explanation for specific data points due to its increased breadth. This quantitative shift was accurately predicted by the Bayesian models.

Discussion

The findings suggest that the human motor system does not merely learn a single mapping of stimulus to response. Instead, it maintains a repertoire of structural models and uses incoming sensory data to weigh which model is currently active. By "integrating out" parameter uncertainty, the brain can make robust decisions about the underlying structure of the environment.

While the study provides a robust proof of concept for Bayesian model selection, future work is needed to disentangle high-level cognitive decision-making from low-level motor circuit logic. Additionally, this paradigm offers a framework for studying how the brain learns entirely new structures (structural learning) rather than just selecting between pre-defined ones.

Summary updated

Abstract

Sensorimotor control is thought to rely on predictive internal models in order to cope efficiently with uncertain environments. Recently, it has been shown that humans not only learn different internal models for different tasks, but that they also extract common structure between tasks. This raises the question of how the motor system selects between different structures or models, when each model can be associated with a range of different task-specific parameters. Here we design a sensorimotor task that requires subjects to compensate visuomotor shifts in a three-dimensional virtual reality setup, where one of the dimensions can be mapped to a model variable and the other dimension to the parameter variable. By introducing probe trials that are neutral in the parameter dimension, we can directly test for model selection. We found that model selection procedures based on Bayesian statistics provided a better explanation for subjects' choice behavior than simple non-probabilistic heuristics. Our experimental design lends itself to the general study of model selection in a sensorimotor context as it allows to separately query model and parameter variables from subjects.
BibTeX
@inproceedings{genewein2012sensorimotor,
  author = {Tim Genewein and Daniel A. Braun},
  title = {{A sensorimotor paradigm for Bayesian model selection}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2012},
  doi = {10.3389/fnhum.2012.00291},
}
Figure