Neural network enhancement method and system and application thereof

Abstract

The invention discloses a neural network enhancement method and system and application thereof. The method comprises the following steps: dividing a given neural network model into a plurality of sub-neural networks; performing initialization processing on each sub-neural network; training each sub-neural network by applying different data enhancement methods, so that each sub-neural network learns different knowledge about the same data; in the training process, calculating a cross entropy loss function of the output of each sub-neural network and a collaborative training loss function between the outputs of each sub-neural network, and combining the cross entropy loss function and the collaborative training loss function to serve as an optimization objective function; and updating the weight parameter of each sub-neural network according to the optimization objective function, and ending the training after a preset number of training times is reached to obtain a trained sub-neural network. According to the method, better image classification performance and performance than a single neural network model can be realized under the condition that the network parameter quantity or the floating point operand is not obviously increased or even reduced.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a neural network enhancement method, system and application thereof. Background technique [0002] Image classification is one of the most fundamental problems in computer vision. Its purpose is to assign a class label to an image, which represents the category of salient objects in the image, such as people, horses, trees, roads, etc. In real life, image classification can be widely applied to scenarios such as smart homes, transportation systems, robots, military, and industrial production lines. In the field of scientific research, the research of image classification is the basis of other computer vision problems, such as object detection, semantic segmentation, pedestrian re-identification, etc. The research of image classification can bring new inspiration and ideas to these fields. In recent years, with the rapid development of convolutional neural networks, deep learning m...

AI Technical Summary

Problems solved by technology

[0007] The technical problem to be solved by the present invention is to provide a neural network enhancement method, system and application for the deficiencies of the prior art, aiming to solve the problem of the model obtained after simply increasing the network width to a certain extent in the prior art. Problem with limited performance gain

Images