Learning in the cerebellum with sparse conjunctions and linear separator algorithms

Investigates potential learning rules in the cerebellum. We review evidence that input to the cerebellum is sparsely expanded by granule cells into a very wide basis vector, and that Purkinje cells learn to compute a linear separation using that basis. We review learning rules employed by existing cerebellar models, and show that results from computational learning theory suggest that the standard delta rule would not be efficient. We suggest that alternative, attribute-efficient learning rules, such as Winnow or incremental delta-bar-delta, are more appropriate for cerebellar modeling, and support this position with results from a computational model