The loss function of sensorimotor learning

Abstract

Motor learning can be defined as changing performance so as to optimize some function of the task, such as accuracy. The measure of accuracy that is optimized is called a loss function and specifies how the CNS rates the relative success or cost of a particular movement outcome. Models of pointing in sensorimotor control and learning usually assume a quadratic loss function in which the mean squared error is minimized. Here we develop a technique for measuring the loss associated with errors. Subjects were required to perform a task while we experimentally controlled the skewness of the distribution of errors they experienced. Based on the change in the subjects' average performance, we infer the loss function. We show that people use a loss function in which the cost increases approximately quadratically with error for small errors and significantly less than quadratically for large errors. The system is thus robust to outliers. This suggests that models of sensorimotor control and learning that have assumed minimizing squared error are a good approximation but tend to penalize large errors excessively.