Adaptive filter models

The original chapter on adaptive-filter models addressed the issue of how far such models were consistent with experimental evidence. Here we consider one particularly important kind of evidence, that concerning the nature of the synaptic plasticity that underlies cerebellar learning. The basic decorrelation learning rule employed by the adaptive-filter model can be translated into spike timing-dependent plasticity form, in which temporal coincidence of parallel fiber and climbing fiber spikes produces LTD at parallel fiber synapses on Purkinje cells, and noncoincidence produces LTP. Although this appears at first sight to be consistent with extensive evidence demonstrating the existence of LTD, other studies have raised serious issues about its functional role. For example, Schonewille et al. (Neuron 70:43–50, 2011) demonstrated that mutant mice unable to sequester Purkinje cell AMPA receptors – the mechanism thought to underlie LTD – did indeed lack LTD, but showed no evidence of impaired learning of typical “cerebellar” tasks such as adaptation of the vestibulo-ocular reflex or eyeblink conditioning. This discrepancy focusses attention on (i) the potential importance of LTP for the initial phases of cerebellar learning, and (ii) the problems posed by reliance on in vitro preparations. An important question concerning the latter is therefore whether the blockage of LTD demonstrated in the mutant mice by Schonewille et al. would also be found in vivo. Resolving this issue would clarify the role of the cerebellum in supervised learning in general, and the plausibility of adaptive-filter models in particular.