What if…? Computational anatomy in the cerebellar microzone, the middle layer of the cerebellar network

Abstract

ABSTRACTIt is generally assumed that the representation of information in the brain is in high resolution and replicable with high fidelity, and that those features make it precise and characterise the inputs to sophisticated computations. Replicability is an important assumption of neural network learning models, including cerebellar models. We question those ideas. The molecular layer of the cerebellar cortex is divided functionally into strips called microzones populated by Purkinje cells and inhibitory interneurons. We propose instead: (1) The microzone computation is a passive effect, unaided, of the triad of cell morphologies, cerebellar network geometry and linear transmission. (2) Control of Purkinje cells is by a chain of rate codes, and not supplanted by learning. (3) Striped topography of input to a network can represent collectively coherent information about body shape and movement. This is capable, passed through the network computation, of driving coordinated motor sequences. (4) A learning algorithm and replicability are unnecessary to explain the evidence.