Research On The Scheme Of Optimizing The Performance Of Classification Model In Weakly Supervised Learning

Machine Deep learning has achieved great success in various fields,such as computer vision,speech recognition and natural language processing,etc.By using massive labeled data sets to train neural network models,it is possible to recognize and classify unknown data.However,in practical scenarios(such as building medical image classifiers),large and high-quality datasets with strong supervision information are usually difficult to obtain due to the highly complex feature of the target data and the insufficient ability and experience of labeling workers.Therefore,weakly supervised learning using datasets with weak supervision information to train models has attracted wide attention of researchers.In weakly supervised learning,the unreliable data set seriously affects the performance of the classification model.For example,the label noise in the training dataset may significantly reduce the performance of classification model.To address the above issue,This thesis systematically investigates how to optimize the performance of classification model under the scenario of weakly supervised learning,i.e.,when facing label noise.The contributions of this thesis are given as follows.(1)Firstly,this thesis introduced the research background and significance of weakly supervised learning,and briefly discussed its basic theory,especially for classification models with label noise.Then,the sources and categories of label noise in samples are summarized.Moreover,From the viewpoints of the three elements of data,model and learning criterion in any classification models,we systematically reviews and summarizes the existing schemes against label noise to optimize the performance of classification models,and analyzes the advantages and disadvantages of these schemes.Finally,according to the current research status,the future research directions in this field were discussed.(2)This thesis proposed a deep neural network classification framework based on calibrating the noisy labels in multi-round(MCDNN),which updates the trained network parameters and calibrates the noisy labels in the current training data in each round of calibration,and the calibrated dataset is used for the next round of training and calibration.Specifically,in each round,the predictive results of some selective anchor samples is utilized to estimate the label transition matrix of the current data,which is then used to infer the weighted average noise rate.Then,through exploiting the "small loss" principle,the noisy data samples are chosen by considering the weighted average noise rate and the trained loss value of data sample.Finally,the noisy labels are calibrated through combining the label transition matrix and predictive result of the trained network adaptively.After multiple rounds of calibration,the noisy level of the dataset is significantly reduced,based on which more accurate classification result can be trained.Experimental results on multiple real datasets show that the proposed method has a large performance improvement compared with existing schemes.(3)Using active learning technology,this thesis proposed an optimization framework for classification model(ALOF),when training dataset contains label noise.Our proposed ALOF framework attempts to select the most valuable samples with noisy label and let them manually relabeled by experts,which are then used to train classification model.In detail,a sample selection strategy combining uncertainty of classification model and sample diversity was designed in this scheme.Specifically,for the target sample,the similarity between it and "anchor samples" of various labels was calculated,and the "extreme anchor samples" with the highest and lowest similarity were found.Then,the labels of the "extreme anchor sample" were compared with the network prediction value of the target sample,and according to the results the target sample was classified as uncertain sample.Finally,the sample set with uncertainty was clustered,and the sample closest to each cluster center was selected as the sample subset that has the greatest value to the model.Compared with the existing selection strategy in several scenarios,the experimental results show that the selection strategy proposed in this thesis has certain advantages.