Hybrid CPU-GPU co-processing scheme for simulating spiking neural networks

Abstract

Many attempts have been made to study the neural networks and to model them. These attempts have led to the development of neural network simulation software packages such as GENESIS [15] and NEURON [18] which have been the de-facto simulators for some time now. However, further studies have found that one of the major hindrances in using the aforementioned simulators is speed. These simulators uses time driven technique which isolates the mimicked time to brief time periods and in every progression the factors of neural states are estimated and reiterated through a numerical examination strategy [9]. This method includes complex calculations which do not foster the development of scalable neural systems. The interest for quick re-enactments of neural systems has offered ascend to the use of alternative reproduction strategy: event driven simulation [12]. The event driven simulation technique just processes and appraises the neural state factors when another event alters the typical advancement of the neuron, that is, the point at which information is created. In the meantime, it is realized that the data communication in neural networks is done by the purported spikes. These occasions are moderately inconsistent and restricted in time. Less than 1% of the neurons are at the same time dynamic [21] and the exercises are amazingly small in numerous apprehensive territories, for example, the cerebellumgranular layer [11][13] which catalyses the efficiency of event-driven Spiking Neural Networks (SNN) simulation. In this work, we present our study on hybrid CPU-GPU based model for simulating SNNs.