Research and implementation of GPU-based parallelization of spiking neural network

Spiking neural networks are sparse, parallel and bionic, which are widely used in deep learning. As spiking neural networks grow, the time cost of training the spiking neural network becomes larger. GPUs, as general-purpose computing hardware, are accessible and suitable for parallel computing, which can effectively accelerate the training of spiking neural network models. However, existing training methods of spiking neural network did not yet make full use of the GPU’s memory, threads, and other resources. We proposed a novel parallelization strategy of training spiking neural network model on GPU to address these issues, including a novel layer-driven simulation strategy based on time-driven and event-driven. This strategy converted the process of serial input into the process of parallel input and avoided a large number of spike-free operations by detecting spikes by layer, while proposing a data pre-caching technique to optimize the transfer and computation time, which effectively improved the computational efficiency of the GPU. Experimental results on the GeForce 2080Ti showed that the proposed strategy was more efficient than the traditional simulation strategy, and the strategy worked well on different datasets, with an absolute speedup ratio of 2.22 on the MNIST dataset and 6.26 on the Caltech Face/Motorbike dataset.