Publication

Abstract : This paper uses scientific computing techniques in understanding cerebellar networks and attempts to model functional behavior of circuits via computational neuroscience. Since highly parallel programmable processors like Graphic Processing Units (GPUs) deliver a high compute capacity at low cost, we modeled granular layer neurons of the rat cerebellum on GPUs and reconstructed a network microcircuit of granular layer for predicting computational properties in such circuits. The main objective of this work was to reconstruct models apt for large scale simulations, involving thousands of neurons while maintaining an acceptable degree of biological details. The hypothesis relating to spatio-temporal information processing in the input layer of the cerebellum has been tested using mathematical modeling. The role of mossy fiber excitation and the modulatory role of Golgi cell inhibition on the granule cells were analyzed. A scalable network consisting of up to 2 million neurons were simulated in millisecond time-scale and the performance efficiency of GPUs over CPUs was compared. The main goal was to understand the scalability issues while implementing such large scale networks and to optimize shared memory access. In GPUs, multi-core parallelization allowed efficient management of computational overheads imposed by synaptic dynamics. We also briefly investigated the performance improvements focusing on decreasing memory access times and the role of optimization techniques. This work is a proof-of-concept implementation apt for densely-packed microcircuits of electrotonically compact neurons with targets to optimize real-time performance and scalability.