Mutual Learning Of Complementary Networks Via Residual Correction For Improving Semi-Supervised Classification

In recent years,with the rapid development of the Internet,a large number of picture data explode increasingly and then how to efficiently process and use these picture data has become a key task for the government,companies and other departments.So,there are three popular ways to deal with image tasks: supervised learning,unsupervised learning and semi-supervised learning.First,people tend to use supervised learning for image classification tasks.However,it inevitably needs to consume a lot of manpower and material resources to label the image data.Later,some researchers in related fields try to use unsupervised learning to process image classification tasks.Although labeled samples were not needed,the results are not satisfactory.Since 1970 s,some scholars used semi-supervised learning methods with only a little labeled data and massive unlabeled data for image classification and then achieved relatively comparable results.Our Work proposes a method called Mutual Learning of Complementary Network via Residual Correction for Improving Semi-supervised Classification.This work mainly explores how to enhance mutual learning between deep convolutional networks for improving semi-supervised classification.Many researches have shown that simply the minimization of the prediction divergence between two separate essential networks may not fully leverage the difference between them.To capture this information,we propose a Complementary Correction Network(CCN),built on top of the essential networks,to correct the prediction of one network,conditioned on the features learnt by another.The resulting more accurate class predictions for the unlabeled instances are used as the training targets to guide the learning of the second essential network so that the second essential network obtains more complementary information than the first one.As a result,our enhanced mutual learning model leads to significant performance gains for image classification,due to the reason that the knowledge learnt by Complementary Correction Network can be ultimately transferred to the first essential network.