A Computational Model of Cerebellar Learning
for Limb Control
Andrew H. Fagg, Nathan Sitkoff, Andrew G. Barto
{fagg | sitkoff | barto}@cs.umass.edu
Computer Science Dept
UMASS-Amherst
James C. Houk
houk@casbah.acns.nwu.edu
Dept of Physiology
Northwestern University School of Medicine
In this paper, we present a model of cerebellar learning for control of limb movements. The model learns to generate relatively smooth and direct arm movements, while not assuming (as do some models) a high-level input that specifies a pre-planned trajectory . The model controls a simulated two-joint arm, which is actuated by a set of six muscles. Control of the arm is shared by a learning cerebellar module, and a hard-wired extra-cerebellar (EC) system. The cerebellar module consists of a group of independent submodules, referred to as Adjustable Pattern Generators (APGs). Each APG is made up of a set of Purkinje cells and a single cerebellar nuclear cell. Activation of an APG's nuclear cell drives a muscle synergy (a subset of the muscles). The EC system is responsible for producing low-quality corrective movements in situations where the cerebellar module is unable to bring the arm to the specified target. Via proprioceptive inputs, the inferior olive assesses directional errors in cerebellar-generated movements by observing subsequent corrective movements produced by the EC system.
For more information, see the poster text.
Presented at 1997 Spring Meeting on the Neural Control of Movement