Heterogeneous Instructive Signals Enable Ensemble Learning in Cerebellar Cortex

The Marr--Albus framework describes learning in Purkinje cells (PCs) through supervised plasticity associated with complex spikes. However, complex spikes occur with low probability in individual PCs following individual sensorimotor events, and PCs are organized into ``microzones'' with shared complex spike tuning and outputs. Here, we unify these observations in a model of associative learning in ensembles of PCs trained under ``heterogeneous plasticity,'' where each PC randomly receives either depression or potentiation during learning. Heterogeneous plasticity critically aids ensemble performance in a pattern recognition task, allowing memorization capacity to scale with ensemble size. An optimal-decoding theory suggests productive roles for recurrent PC--PC dynamics and nonlinear dendritic integration in the cerebellar nuclei. We account for several additional experimental observations, including bursting and pausing PCs, a bias toward ``upbound'' or ``downbound'' learning mechanisms, and correlations in PC physiology and synaptic weights. Together, our results suggest that cerebellar microzones subserve an ensemble learning computation enabled by heterogeneous plasticity and nonlinear information processing.