Learning of Motor Sequences Based on a Computational Model of the Cerebellum

Abstract

In classical conditioning, the repeated presentation of a Conditioning Stimulus (CS) followed by an Unconditioned Stimulus (US) establishes a basic form of associative memory. After several paired CS-US presentations, a Conditioned Response (CR) is elicited by the solely presence of the CS. It is widely agreed that this associative memory is stored in the cerebellum. However, no studies have link this basic form of cerebellar associative learning with the acquisition of sequences of motor actions. The present work suggests that through the Nucleo Pontine Projections (NPPs), a CR elicited by a first CS may be fed-back to the cerebellum, and that this CR can act as the CS for a subsequent CR. This process would allow a single CS to trigger a sequence of learned responses, having a total duration above the timespan of the cerebellar memory trace. We demonstrate this principle with a robotic experiment, where a computational model of the cerebellum that includes the NPPs controls a robot navigating a track with two turns. A predictive cue, the CS, precedes the first turn, but the second one can only be acquired if the previous turn is also used as a CS. After repeated training trials, the robot associates a sequence of two turns to the single CS. This result confirms that the positive feedback established via the NPPs allows the cerbellar model to control an action sequence, and that the duration of the whole sequence can exceed the timespan of a cerebellar memory trace.