Image recognition-oriented neural network training method and system

Abstract

The invention provides a neural network training method and system for image recognition. The method comprises the following steps: S1, acquiring a training image data set; s2, constructing a neural network model, and setting a loss function as J < theta; x, y) is a preset loss function, and g (w) is a regularization item; s3, inputting training data in a training image data set into the neural network model for forward propagation to obtain an image feature vector; s4, calculating a loss function value, and judging whether the neural network model is converged or not; if yes, training is ended, and if not, the step S5 is executed; s5, according to the loss function calculation gradient, back propagation is carried out on the neural network model, the weight of each layer of the neural network model is updated, and the weight of each layer is updated according to the gradient direction for reducing the loss function value; and S6, progressively increasing the number of iterations, and returning to the step S3. According to the method, the neural network with more optimized structure and stronger robustness generalization can be obtained.

Description

Technical field [0001] The present invention relates to the technical field of computer software, and more particularly to a neural network training method and system for image recognition. Background technique [0002] Depth neural networks have been successful in many applications, but some models can have good generalization ability to extensive application scenarios under limited size data sets. It is still a mystery. In order to solve this problem, we need to better understand the reasons for the generalization of deep learning models, which can also bring a lot of benefits, such as providing safety for safe critical application scenarios (such as automatic driving / financial industry). Guaranteed to design a better depth neural network model, further enhance the impliance of the network. [0003] There are currently two aspects of the research field of neural network generalization: [0004] 1. Existing work Most focuses on improving the accuracy of the neural network in t...

AI Technical Summary

Problems solved by technology

[0006] Based on the above problems, the generalization error of the current neural network is still large, and the robustness of network generalization is poor. When it is applied to safety-critical fields such as autonomous driving and intelligent medical care, there are greater risks and hidden dangers. It is very likely to cause huge human and economic losses

Images