Related Experiment Videos
The loss function of sensorimotor learning
Konrad Paul Körding1, Daniel M Wolpert
1Sobell Department of Motor Neuroscience, Institute of Neurology, University College London, Queen Square, London WC1N 3BG, United Kingdom. konrad@koerding.de
Summary
People optimize motor learning using a loss function that penalizes small errors quadratically but large errors less severely. This makes the system robust to outliers, unlike models that excessively penalize large errors.
Area of Science:
- Neuroscience
- Motor Control
- Human Performance
Background:
- Motor learning involves optimizing performance based on a task-specific loss function.
- Current models often assume a quadratic loss function, minimizing mean squared error.
- This assumption may inaccurately penalize large movement errors.
Purpose of the Study:
- To develop a novel technique for measuring the loss function associated with motor errors.
- To experimentally investigate the human loss function during a motor task.
- To compare inferred human loss functions with standard quadratic models.
Main Methods:
- Subjects performed a motor task with experimentally manipulated error distribution skewness.
- Inferred the underlying loss function from changes in subjects' average performance.
- Utilized a novel error-measurement technique to quantify loss.
Main Results:
- Human motor learning employs a loss function with approximately quadratic cost for small errors.
- The cost increases significantly less than quadratically for large errors.
- This indicates a built-in robustness to outliers in the human motor system.
Conclusions:
- Human sensorimotor control is robust to large errors, deviating from a strict quadratic loss.
- Standard models assuming mean squared error minimization may excessively penalize outliers.
- Findings suggest a more nuanced understanding of loss functions in motor learning is needed.