How cerebellar architecture facilitates rapid online learning

Abstract

SO_SCPLOWUMMARYC_SCPLOWThe cerebellum has a distinctive architecture in which inputs undergo a massive size expansion in dimensionality in early layers. Marr and Albuss classic codon theory and more recent extensions1-4 argue that this architecture facilitates learning via pattern separation. The essence of this idea is that sparsely active clusters -- codons-- of inputs are more easily separable in a higher dimensional representation. However, recent physiological data indicate that cerebellar activity is not sparse in the way anticipated by codon theory. Moreover, there is a conceptual gap between static pattern separation and the critical role of the cerebellum in dynamic tasks such as motor learning. We use mathematical analysis and simulations of cerebellar learning to identify specific difficulties inherent to online learning of dynamic tasks. We find that size expansions directly mitigate these difficulties, and that this benefit is maximised when granule cell activity is dense.