The uniform structure of the cerebellar microcircuit has long suggested to theoreticians that it implements a powerful signal-processing algorithm, that can be applied (via appropriate external connections) to a wide variety of sensorimotor tasks.

Current computational attempts to identify this algorithm appear to be converging on the adaptive-filter model of the cerebellar microcircuit, first proposed by Fujita in 1982. One reason for this choice is the broad resemblance between the structure of an adaptive-filter and the wiring of the cerebellar microcircuit. A second is that in control and engineering contexts adaptive-filters can carry out the sorts of computational task associated with cerebellar function, including the removal of interference from sensory signals and the generation of accurate motor commands. Thirdly, they do so in an intuitively satisfying way, by decorrelating their main (mossy fibre) inputs from a (climbing fibre) teaching signal, using a learning rule that appears to have a direct physiological counterpart.

These arguments suggest that the adaptive filter model can be taken as a serious hypothesis of cerebellar microcircuit function, thereby raising important questions about exactly how it can be tested. A key issue is how an abstract signal-processing algorithm can be implemented by populations of slow, noisy, spiking neurons. A better understanding of this issue will enable us to determine whether certain experimental findings, such as Purkinje cell bistability or olivary synchrony, are actually compatible with the adaptive-filter algorithm, or are evidence for genuinely new computational principles of cerebellar function.

Finally, the adaptive-filter hypothesis focuses attention on the precise connectivity of cerebellar modules, particularly their outputs and climbing fibre inputs. Only in very few cases are these well understood. Their further investigation will cast light not only on cerebellar function itself, but also on the general computational question how overall intelligent behaviour can constructed from simpler components.

 

Video registration (1h26min)