Pulse neural network simulation method based on GPU

Abstract

The invention discloses a GPU-based spiking neural network simulation method. The method comprises the steps of initializing a neural network structure and a network weight; loading a data set and carrying out pulse coding by adopting a pulse coding module; calling a GPU calculation module to select a proper GPU to calculate the pulse neuron membrane voltage according to the data volume, the calculation amount and the priority of the calculation task, comparing whether the pulse neuron membrane voltage exceeds a threshold value or not, and issuing a pulse; creating a pulse queue, and adding neurons of a trigger pulse into the pulse queue; if the pulse queue is not empty, finding out a post-synaptic neuron corresponding to the next layer according to the network structure, and repeating the steps S3-S5 until an output layer is reached; and calculating a loss function according to the result of the output layer and the actual pulse result, and updating the neural network by adopting a gradient descent mode until the iteration is completed. According to the invention, the training speed of the spiking neural network is accelerated, the advantages of the spiking neural network in the aspects of low power consumption and low time delay are exerted, and the conditions of overlarge data set, insufficient video memory and incapability of training are avoided.

Description

technical field [0001] The invention belongs to the technical fields of pulse neural network, brain-inspired computing and high-performance computing, and in particular relates to a GPU-based pulse neural network simulation method. Background technique [0002] The spiking neural network is the third generation of artificial neural network, which uses spiking neurons as the basic unit and is connected to each other to form a neural network. The membrane potential of a spiking neuron accumulates in a non-linear manner, and when the threshold potential is reached, a pulse is released and enters the refractory period to cool down, while the pulse is transmitted to the next neuron. Large-scale spiking neural networks can not only obtain powerful computing power, but also better verify biodynamic properties. [0003] With the maturity of spiking neural network technology, spiking neural networks tend to be deeper and larger networks. The method of using hardware to accelerate s...

AI Technical Summary

Problems solved by technology

[0004] However, the existing GPU-based spiking neural network simulation technology still has the following disadvantages: most of the spiking neural network models are time-driven, and do not give full play to the advantages of spiking sparsity; most GPU-based spiking neural networks are accelerated by GPU , cannot be extended to multi-nodes for acceleration; there are a large number of invalid operations in the GPU pulse neural network training, wasting computing resources

Images