Hardware friendly pulse neural network model based on STDP non-supervised learning algorithm

Abstract

The invention discloses a hardware friendly pulse neural network model based on an STDP non-supervised learning algorithm. The pulse neural network model is a polysynaptic time-delay forward feedback neural network model, each-layer neural elements are in interconnection with neural elements of other layers through multiple synapses, and internal neural elements of each layer are mutually independent. The model is advantaged in that model design of the pulse neural network is carried out completely based on a digital hardware circuit platform, the STDP non-supervised learning algorithm is innovatively applied to learning training of the pulse neural network model, special functions can be realized, the hardware platform pulse neural network model has relatively high stability, a relatively fast speed is realized on the condition that occupied hardware resources are relatively a few, moreover, a network structure has a parallel connection characteristic, expansibility is substantially improved, and a new concept is provided for realizing a super-scale pulse neural network.

Description

technical field [0001] The invention belongs to the pulse neural network technology, in particular to a hardware-friendly pulse neural network model based on the STDP non-supervised learning algorithm. Background technique [0002] In the biological brain, hundreds of millions of neurons are connected to each other through synapses for information transmission, thereby manipulating individuals to carry out various complex and meticulous biological activities. This kind of nervous system has been studied and imitated by people, and artificial neural networks have also been attention worldwide. In recent years, as the third generation of artificial neural network, spiking neural network has been widely studied due to its amazing biological similarity and powerful computing ability in pattern recognition, image processing, computer vision and other aspects. So far, many scholars have focused on various spiking neural network models and learning mechanisms. These computational ...

AI Technical Summary

Problems solved by technology

However, most of the current work is concentrated on the computer software platform. On the one hand, due to the summary of previous work, the calculation model of the software platform on the spiking neural network is relatively mature. On the other hand, the development difficulty of the software platform from the learning algorithm to the entire neural network system The hardware platform is much simpler than the hardware platform, but the serial execution characteristics of the software platform fundamentally do not conform to the parallel processing characteristics of biological brain neurons, and the problems of long simulation time and poor scalability are also prominent.

Images