BP neural network learning circuit

Abstract

The invention discloses a BP neural network learning circuit, mainly comprising a differencing circuit module, a summing circuit module, a weight storage circuit module and a weight refreshing circuitmodule, wherein the differencing circuit is mainly used for calculating an error; the summing circuit is used for completing a weight voltage accumulation function; the weight storage circuit is usedfor storing current weight voltage at two ends of a capacitor; and the weight refreshing circuit is used for refreshing the current weight voltage by charging and discharging the capacitor. The mainpurpose of the BP neural network learning circuit is to realize the function that the weight in the BP neural network can be continuously adjusted. Moreover, the number of times of learning can be controlled through the pulse switch, so that the learning proportion of each sample can be the same under the control of the pulse switch under the condition of learning of a plurality of samples, and thus, when the learning circuit is applied to the BP neural network, the BP neural network learning system can learn the characteristics of each sample indiscriminately.

Description

technical field [0001] The invention belongs to the field of BP neural network, relates to a BP neural network realized based on a hardware circuit, and in particular relates to a learning circuit. Background technique [0002] Since the advent of computers, various complex scientific calculations and information processing do not require the human brain to complete, because in many complex calculations, if a clear calculation method is given, the computer can quickly calculate the results, while the human The calculation speed of the brain is far less than one ten-thousandth of it; however, when some information is not clear or the information given is not very accurate, the computer seems to be somewhat powerless, especially for some problems that require reasoning. It is greatly restricted, so there is the development of artificial neural networks. As early as 1943, the neural network mathematical model was proposed by psychology professor Warren McCulloch and mathematic...

AI Technical Summary

Problems solved by technology

[0003] However, the traditional learning algorithm of BP neural network, that is, the gradient descent method, has some obvious disadvantages, for example, it is easy to fall into a local minimum during the learning process, or the convergence speed is very slow when the input is small or the input data is at both ends of the activation function. Slow, so at present, most researchers at home and abroad are studying the improvement of the BP neural network algorithm, or the improvement of the traditional algorithm itself, such as adding momentum method, adaptive learning rate method and introducing steepness factor, etc. Or change the algorithm, such as using Newton's method, LMBP algorithm, genetic algorithm, etc.

Images