Adaptable internal representations drive cerebellum-mediated predictive control of an innate behavior

Abstract

The cerebellum is known to optimize behavioral responses via error minimization, yet how past sensory experience shapes cerebellar output is still poorly understood. We imaged cerebellar Purkinje cells (PCs) in larval zebrafish during a behavioral assay in which they learnt to expect the direction of optic flow stimulus. PCs encoded signals of expectation and a graded error signal when a deviant stimulus was encountered. The amplitude of the error signal scaled based on stimulus history over minute-timescales. Granule cells (GC) did not encode error, however, some GC types encoded putative expectation signals. When PC population-wide readouts of expectation and error signals were strong, swim latency was low and vice versa, indicating participation of these signals in motor planning. Consistent with this, lesioning the cerebellum abolished expectation-dependent modulation of swim latency. Based on these results we propose that PCs use expectation and error signals to acquire and update internal models of the world. HighlightsO_LILarval zebrafish learn to expect repeatedly encountered optic flow stimuli C_LIO_LIStimulus history over minutes contributes to an internal model of the sensory world C_LIO_LIPurkinje neuron activity encodes deviations from expected direction of optic flow C_LIO_LIThe cerebellum uses learnt expectations to generate lower latency swim responses C_LI