Simulation and training method of pulse neural network with pulse moment offset

Abstract

The invention discloses a simulation and training method of a pulse neural network with pulse moment offset, which comprises the following steps of: discretizing to update a neuron state quantity step by step, estimating whether pulse distribution is generated in a time step range and the moment offset thereof, calculating pulse input at the next moment according to synaptic connection and repeating the process to complete network simulation; and according to simulation result delay neuron state quantity, pulse emission quantity and time offset back propagation gradient error, updating network parameters. Errors in time step-by-step simulation calculation of the pulse neural network are reduced, and the simulation efficiency is improved under the same precision condition. And meanwhile, the pulse error is transmitted through two dimensions of the pulse emission amount and the time offset in the error back propagation algorithm, so that the training efficiency is improved.

Description

technical field [0001] The invention relates to the fields of neural network technology and brain-inspired computing technology, in particular to a simulation and training method of a pulse neural network with a pulse time offset. Background technique [0002] Spiking neural network (SNN) is a neural network technology that more accurately simulates the calculation and communication mechanism of the biological nervous system and is closer to the information processing method of the human brain. The dynamic mechanism of pulse generation of spiking neurons makes SNN have inherent ability to process historical information. Compared with traditional artificial neural networks, it has unique advantages in processing spatio-temporal information. At the same time, the way of transmitting information by pulses saves computing resources. Therefore, the spiking neural network is an important technical tool in the field of brain-inspired intelligence. [0003] However, the continuous ...

AI Technical Summary

Problems solved by technology

This clock-driven calculation method causes neurons to only calculate the pulse emission at the time step, and there is a certain error between the pulse time based on the time step and the real pulse time.

Through pulse transmission, the error existing at the pulse time will be further enlarged, which seriously affects the simulation calculation accuracy of SNN

In order to ensure the calculation accuracy, a smaller time step is usually required, but a smaller time step increases the calculation amount of simulation and training, and reduces the calculation efficiency

Secondly, because SNN has non-differentiable mutations in the process of pulse generation, SNN cannot directly use the standard gradient error backpropagation method for learning

Existing SNN algorithms usually use the method of substituting gradients for pulse transmission to reverse the error through the state quantity at each time step, but the existing substituting gradient methods cannot approximate the gradient changes in the two directions of pulse emission probability and pulse time movement, so the SNN network There is a large deviation in the gradient calculation of

In summary, the existing SNN clock-driven calculation method causes large errors in SNN simulation and training, which limits the application of SNN technology in complex artificial intelligence tasks

Images