Research On Inaccurate Supervision Learning And Its Application

Supervised learning algorithm has achieved great success in theory and application,but the effect of such algorithm dependent on the sample label quality of training data set,in the actual problem get training samples with high quality labels are often time-consuming.in theory and application In order to save human and material resources,manpower,web crawler,crowdsourcing method for alternative methods used for acquisition of training data.However,the data obtained by these methods often have a large number of false labels,namely,label noise,which will bring a lot of negative effects on the network training.Therefore,the research on the processing method of label noise has the greatest significance in promoting the application of machine learning engineering and reducing the cost of machine learning.This paper focuses on the classification of images with label noise.The main research contents are as follows:(1)Weakly supervised learning methods based on curriculum learning is studied.Firstly,the method of ranking the samples with the probability of having accurate labels by using k NN algorithm,proposed the curriculum learning based on k NN algorithm.Then,the method of initialization center point in K-means clustering algorithm is improves,and a method of curriculum learning based on improved K-means clustering algorithm is proposed.(2)Weakly supervised learning methods based on joint optimization of metwork parameters and sample labels is studied.First,the joint optimization of network parameters and sample labels is proposed.Then,the main steps of the joint optimization based on network parameters and hard labels and the joint optimization based on network parameters and soft labels and the setting of loss function are studied.(3)In the neonatal pain expression database and cifar-10 data set,the weakly supervised learning method based on curriculum learning and based on joint optimization of metwork parameters and sample labels are evaluated respectively.The experimental results show that the best accuracy rate of the curriculum learning method based on KNN algorithm is 78.5%.On cifar-10 data set,the joint optimization based on network parameters and soft labels achieves the best test accuracy in random classification label noise and class-conditional label noise.