Impulse neural network reward optimization method and device, electronic equipment and storage medium

Abstract

The invention provides a spiking neural network reward optimization method and device, electronic equipment and a storage medium. The method comprises the steps of performing an initialization operation on a spiking neural network; based on a pulse encoder, converting the sample data into a sample pulse sequence code; and inputting the sample pulse sequence codes into the pulse neural network, determining the distribution state of each layer of neurons in the forward propagation process of the pulse neural network, then determining the reward corresponding to each hidden layer, and outputting the hidden layer based on the output of each hidden layer and the corresponding reward. optimizing the synaptic weight between each layer of neurons and the corresponding pre-synaptic neurons until the spiking neural network converges; wherein the optimization operation of the synaptic weight between each layer of neurons and the corresponding pre-synaptic neurons is mutually independent. The invention has biological rationality, and compared with a layer-by-layer optimization mode, unnecessary information storage can be reduced, memory occupation is reduced, energy consumption is reduced, and the method can be easily placed on a chip for use.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, and in particular, to a reward optimization method, device, electronic device and storage medium of an impulse neural network. Background technique [0002] The spiking neural network is considered to be the third generation of artificial neural network. The basic unit of information transmitted between neurons in a spiking neural network is a discrete pulse containing the precise time at which the membrane potential state reaches the firing threshold. This event-type signal contains internal neuronal dynamics and historically accumulated (and decayed) membrane potentials. Spike training in spiking neural networks opens up a new temporal coordinate for better characterization of processing sequence information compared to firing rates in artificial neural networks (where the firing rate can be defined as an analog value describing the propagating information). In addition to neu...

AI Technical Summary

Problems solved by technology

[0004] The present invention provides a reward optimization method, device, electronic equipment and storage medium for a pulse neural network, which is used to solve the problem of using the BP training method to train the pulse neural network in the prior art, which does not meet the biological learning method criteria, and applying the trained pulse neural network The defect that the recognition effect is not good when it comes to the recognition task

Images