Conversion method of high-precision low-delay pulse neural network

Abstract

The invention discloses a conversion method of a high-precision low-delay pulse neural network, which comprises the following steps: constructing an ANN with the same structure as an SNN and adding a BN layer, then training the ANN, folding the BN layer into weight and deviation and modifying the network structure, then copying the weight of the ANN into a layer corresponding to the SNN, changing an activation function into an IF model and setting simulation time, finding the optimal threshold value of each layer in the SNN network layer by layer in a self-adaptive threshold value mode, finally, calculating and comparing the output of the target SNN network and the output of the ANN network, carrying out deviation correction, membrane potential correction and weight correction on the target SNN network layer by layer, and obtaining the high-precision low-delay pulse neural network. According to the method, the optimal threshold value under the current simulation time is selected, and the parameters of the SNN are corrected layer by layer, so that the performance of the converted SNN can be very close to that of the ANN, and very high performance can be obtained under the condition that the simulation time is relatively short, namely, the SNN with low delay and high performance is realized.

Description

technical field [0001] The invention relates to the technical field of pulse neural networks, in particular to a conversion method for obtaining high-precision and low-delay pulse neural networks through layer-by-layer fine-tuning and self-adaptive thresholds. Background technique [0002] Artificial neural networks 1 (Artificial Neural Networks, ANN) is the most important achievement that people have obtained in the study of brain neurobiology. It has made breakthroughs in many tasks and achieved performance beyond human beings. It is called the second-generation neural network. The current artificial neural network transmits information from front to back layer by layer through ultra-high-precision values, and makes judgments by statistically outputting the size of the output layer, but this method is very ideal and has nothing to do with the working mechanism of real biological networks. The Spike Neural Networks (SNN) is called the third-generation neural network model,...

AI Technical Summary

Problems solved by technology

[0016] Aiming at the above-mentioned technical problems, the present invention provides a conversion method for obtaining high-precision and low-delay impulse neural network through layer-by-layer fine-tuning and self-adaptive threshold, so as to solve the problem that the target SNN is in a very short simulation time after converting from the source ANN to the target SNN. Insufficient network performance

Images