SNN training method and device, storage medium, chip and electronic device

Abstract

The invention discloses an SNN training method and device, a storage medium, a chip and an electronic device. In order to solve the technical problems of instability, low efficiency and poor precision when the spiking neural network is trained, a time constant or an attenuation coefficient in the spiking neural network is trained through a space domain and time domain error return allocation strategy. The technical scheme disclosed by the invention not only considers the spatial domain dependency of the pulse neural network, but also considers the time domain dependency between pulses, so that the technical problem of how to efficiently and stably train the time constant or the attenuation coefficient in the space-time domain is solved; the technical effects that the SNN can be trained more quickly and stably, and the brain-like chip deployed with the SNN configuration parameters obtained through training has higher reasoning precision performance are achieved. The method is suitable for the fields of pulse neural network training and brain-like chips.

Description

technical field [0001] The invention relates to an SNN training method and device, a storage medium, a chip and an electronic device, and in particular to a training method and device, a storage medium, Chips and Electronic Devices. Background technique [0002] The configuration parameters of a spiking neural network (SNN) usually include parameters related to spiking neurons such as synaptic weights, time constants (corresponding to the decay coefficient threshold), etc. In the prior art, when training a spiking neural network, only the influence of a single time step is usually considered, and the time domain dependence between the pulses is ignored. To this end, prior art 1 proposes a time-domain return assignment strategy for back-propagating errors to the front layers of the network. However, the applicant found that the solution is unstable, the convergence speed is slow, and the inference accuracy of SNN is not high. [0003] For this reason, the applicant propose...

AI Technical Summary

Problems solved by technology

However, the applicant found that the scheme was unstable, the convergence speed was slow, and the reasoning accuracy of SNN was not high.

[0003] For this reason, the applicant proposed a space-time domain spiking neural network training method in prior art 2, which overcomes this technical problem relatively well, but this training method is only for configuration parameters of the type of synapse weight, It is not yet known how to train in the space-time domain for configuration parameters such as time constants

Images