Cerebellar Model Controller with new Model of Granule Cell-golgi Cell Building Blocks and Two-phase Learning Acquires Multitude of Generalization Capabilities in Controlling Robot Joint without Exponential Growth in Complexity

Abstract

Processing in the cerebellum is roughly described as feed forward processing of incoming information over three layers of the cerebellar cortex that send intermediate output to deep cerebellar nuclei, the only output from the cerebellum. Beside this main picture there are several feedback routes, mainly not included in models. In this paper we use new model for neuronal circuit of the cerebellar granule cell layer, as collection of idealized granule cell–golgi cell building blocks with capability of generating multi-dimensional receptive fields modulated by separate input coming to lower dendrite tree of Golgi cell. Resulting cerebellar model controller with two-phase learning will acquire multitude of generalization capabilities when used as robot joint controller. This will usually require more than one Purkinje cell per output. Functionality of granule cell-Golgi cell building block was evaluated with simulations using Simulink single compartment spiking neuronal model. Trained averaging cerebellar model controller attains very good tracking results for wide range of unlearned slower and faster trajectories, with additional improvements by relearning at faster trajectories. Inclusion of new dynamical effects to the controller results with linear growth in complexity for inputs targeting lower dendrite tree of Golgi cell, important for control applications in robotics, but not only.