How rate information is collectively coded in input to the cerebellum: a hypothesis and simulation

Abstract

ABSTRACTIn the cerebellum, below a scale threshold, cells seem to select their targets at random and receive contact from a random sample of cells that innervate their location. The largest source of glutamatergic input to the cerebellum is from mossy fibres, which terminate on granule cells in the inner layer of the cerebellar cortex, the granular layer. The cerebellar cortex is divided functionally into long thin strips called microzones. Information received by a strip is mixed. We hypothesise that mossy fibre rate information is contained in the collective activity of cells that innervate a region the size of a microzone, and that random connectivity is a computational strategy. The physiology of the granular layer is adapted and essential to implement the strategy. Information is encoded in statistical properties of firing rates that are received as input to a whole strip. There is no finer grade of resolution notwithstanding the substantial data loss. Thousands of mossy fibres innervate a strip. However, as the code is contained in rate statistics, it is not necessary to receive all signals to ‘read’ it. It is only necessary to receive a statistically representative random sample. This has the consequence that the output of recoding (granule cell signals), is functionally unitary at the same scale (a whole strip). To test these ideas, we use the Kruskal-Wallis and Bayes factor tests and biologically-detailed computer simulations. The main finding is that group code in the form described is computationally workable and physiologically plausible and can explain the evidence.SIGNIFICANCE STATEMENTIt is an almost universal theoretical unknown if and to what extent, and how, brain regions work blind, i.e., unable to ‘see’ upstream. We propose that the cerebellar solution is to exploit it, treating information as homogeneous on a surprisingly large scale. This runs counter to typical theoretical thinking because there is a large loss of data. Nonetheless, it is able to explain detailed physiological evidence and has the large biological advantage that it simplifies wiring (and genetic code). Collective neural code is likely to be a widespread feature of the way the brain represents information. Random sampling below a scale threshold, within topographically defined regions, provides a mechanism to read it.