Accelerating Allen Brain Institute’s Large-Scale Computational Model of Mice Primary Visual Cortex

Abstract

AbstractEfficient simulation of large-scale biological neural networks is important for studying the working mechanisms of the brain. Allen Brain Institute proposed a large scale computational model of mice primary visual cortex (V1) which is so far the most detailed model of mice V1. The original model is composed of two parts: lateral geniculate nucleus (LGN) model and V1 spiking neural network (SNN) model. The original model has low computational efficiency and does not support GPU and multithreading acceleration. In this work we we propose several techniques for accelerating the original model. We refactored the original LGN model based on PyTorch and V1 SNN model based on NEST-Simulator, and enabled GPU acceleration and multithreading and multiprocess acceleration. Our LGN model achieved 60 times acceleration in computing speed. The building time of V1 SNN model was accelerated by 5.7 times. When using multiple threads and process, our V1 SNN model achieved 17.8 times acceleration on clusters compared with the original model. Our refactored model is helpful for computational research about the mice V1.KeywordsSpiking neural networkPrimary visual cortexComputational acceleration