Fusion neuron model, neural network structure, training method, reasoning method, storage medium and equipment

Abstract

The invention relates to an artificial neuron and a neural network, in particular to a fusion neuron model, a neural network structure, an inference method and a training method of the neural networkstructure, a computer readable storage medium and computer equipment. Each synaptic connection weight of the fusion neuron model is any continuous derivable nonlinear function, and linear-to-nonlinearmapping is realized on the synaptic weight. The neural network structure forms a hierarchical structure by taking a fusion neuron model as a basic composition unit, and the reasoning method comprisesthe following steps: substituting input data into a connected nonlinear weight function, calculating a connection weighting result, and summing all weighting results of neurons, and directly transmitting the weighted results to a next-stage neuron; carrying out forward transmission in sequence to finally obtain an identification result. According to the training method, parameters of a neuron model are optimized through a back propagation algorithm and a gradient descent algorithm; and the computer readable storage medium and the computer equipment can realize specific steps of the reasoningmethod and the training method.

Description

Technical field [0001] The present invention relates to artificial neurons and neural networks, specific to a fusion neuron model, a neural network structure, a method of reasoning, a training method, a computer readable storage medium, and a computer device. Background technique [0002] Under the promotion of the short tide of the new technology revolution, intelligent has become an inevitable trend of future social form, and artificial intelligence technology plays an increasingly important role in the information era. Data processing technology with artificial neural networks is a mainstream method of today's artificial intelligence, which interprets data with a mechanism of anticipants, and forms more abstract high-level properties by combining low-level features. Currently, artificial neural network technology has extensive applications in pattern identification, image processing, intelligent control, combined optimization, financial prediction, communication, robots, and e...

AI Technical Summary

Problems solved by technology

[0006] The present invention solves the problem that when analog computing devices are used to realize traditional artificial neurons and network computing models, there are a large number of required analog device combinations, which are prone to drift due to environmental interference, and part of the activation functions are difficult to use analog de

Images