Hybrid particle swarm pulse neural network mapping method for power consumption

Abstract

The invention discloses a hybrid particle swarm pulse neural network mapping method for power consumption. On the basis of a neuron node mapping way with combination of a particle swarm algorithm and a genetic algorithm, an optimal mapping result of mapping of a neuron node to a hardware system is obtained. The mapping method employs the particle swarm algorithm; the mutation operation of the genetic algorithm is combined during the particle swarm algorithm operation process and thus the basic particle swarm algorithm is improved; and the algorithm is circulated until a terminal condition is met. Therefore, defects of performances of the basic particle swarm algorithm can be overcome and the original searching capability of the particle swarm algorithm is utilized completely; a defect that the basic particle swarm algorithm is convergent too early and thus is easy to fall into local optimal node can be overcome; and thus the algorithm can search a global optimal solution. And thus the power consumption of the system can be reduced and the applicability of the mapping plan can be enhanced.

Description

technical field [0001] The invention relates to the field of intelligent optimization, in particular to a power consumption-oriented hybrid particle swarm impulse neural network mapping method. Background technique [0002] The research of Spiking Neuron Networks (SNN) has increasingly become a research hotspot in the field of computational intelligence. The spiking neural network adopts a coding method based on time spike sequences, which is closer to the understanding of the biological nervous system in brain science. Compared with the traditional neural network, the spiking neural network shows stronger bionic characteristics and computing power. [0003] As the most biologically realistic artificial neural network model so far, all neurons of the spiking neural network have a potential pulse trigger mechanism similar to that of biological neurons. This mechanism makes the spiking neural network different from the traditional artificial neural network based on pulse freq...

AI Technical Summary

Problems solved by technology

But mammalian brains have more than 10 neurons 10 Therefore, the scale of the spiking neural network must be very large, and the existing methods cannot effectively provide interconnections between millions of neurons / synapses

The use of conventional software computing methods to simulate spiking neural networks, such as pure software simulation based on traditional computers, cannot fully reflect the parallelism advantages of spiking neural networks, and its execution speed is too slow to be real-time and cannot be simulated on a large scale. The spiking neural network performs calculations, and the scalability of the system is poor

However, common hardware implementations, such as parallel GPUs, have disadvantages such as high power consumption

Images