Noisy Data Classification Based On Deep Learning

Although deep models achieve excellent generalization performance in many applications,these deep models rely heavily on accurately labeled large-scale training sets.In the case of label noise in the training set,the generalization performance of these deep models is seriously threatened and challenged.This dissertation proposes several methods for label noise learning,such as designing a robust loss function for multi-classification,indicating the convergence path of deep model in the presence of label noise,proposing a dynamic label learning algorithm to deal with label noise,and complementary label learning to address label noise.The main research contributions and innovations of this dissertation are as follows:(1)To address the problem of how to use robust binary losses for multi-classification under label noise,this dissertation adopts the multi-category large margin classification strategy to construct the objective loss for a multi-classification model.That is,the output of the label is taken as the positive class,and the other outputs are taken as the negative class to construct the objective loss.This objective loss can be completed by using Pairwise-Comparison(PC)or One-Versus-All(OVA)to employ robust binary loss.This dissertation also theoretically demonstrate that our method is inherently tolerant to label noise.Experimental results show that the proposed method is robust to label noise and exceeds the current advanced methods.(2)In view of the opaque learning process of deep models under label noise,this dissertation proves that any loss function can learn the optimal model corresponding to the noisy labels through in-depth theoretical analysis and derivation.In addition,this dissertation shows a general convergence path to deep models under noisy labels.Furthermore,a label-corrected algorithm is designed based on the theorems provided in this dissertation so that any loss function can learn the corresponding optimal model.Experimental results show that this approach has obvious advantages compared to other methods.(3)This dissertation creatively designs a dynamic label learning algorithm to train deep models under label noise to solve the problem that deep models are prone to overfit the noisy ones.More importantly,the algorithm can use abundant loss functions designed for traditional deep learning to learn from label noises,even the cross-entropy loss function,which has been proved as not robust to label noise under the traditional deep learning algorithm.This dissertation proves that the proposed dynamic label algorithm is robust to label noise strongly.Moreover,this algorithm is independent of loss function and label distribution.In addition,this dissertation theoretically analyzes the convergence process of the dynamic label learning algorithm and the performance of the model under the noisy training set,and clearly explains the robustness and classifier consistency of the algorithm against label noise.Experimental results not only verify the theoretical analysis of this method,but also show that the dynamic label learning algorithm designed in this dissertation exceeds other advanced learning algorithms.At the same time,results shows that the algorithm has strong robustness,expansibility,and versatility.(4)Current studies show that the problem of label noise learning can be solved by complementary label learning.However,the problem of complementary label learning is mainly solved by designing specific objective loss functions.To address the problem that many loss functions designed for ordinary label learning cannot be effectively used for complementary label learning,this dissertation profoundly studies the characteristics of complementary labels as well as the existing theoretical results,designs a general complementary label learning objective function with theoretical proof that any loss function can be used for complementary label learning under this objective function,and points out how to use the corresponding verification set to select the optimal model.Experimental results thoroughly verify the correctness of the theoretical analysis and the improvement of our proposed method over current state-of-the-art methods.(5)Aiming at the problem of how to ensure classifier consistency between complementary label learning and ordinary label learning,this dissertation first designs a general objective loss function for complementary label learning,which enables the use of many loss functions designed for ordinary label learning to train the model under complementary labels.Two sufficient conditions on a loss function are given to guarantee the classifier consistency,and a general form of loss function that does not have classifier consistency is also provided in this dissertation.In addition,the difference in generalization performance of different loss functions in complementary label learning is analyzed in depth.Experimental results show that the proposed method achieves better generalization performance than the most advanced complementary label learning algorithms,and the theoretical explanation given in this dissertation is fully verified.The research results have important theoretical value and practical value.In this dissertation,the robustness of the symmetric loss function to label noise is studied based on the objective function.In addition,the general convergence path of deep model learning under label noise is further studied.A dynamic label learning algorithm is designed to make the most of loss functions robust to label noise.For the new learning paradigm,complementary label learning,this dissertation designs an objective function framework that extends the loss function of original label learning to complementary label learning and theoretically shows that the validation set can be used to select the optimal model.Lastly,an objective loss function with classifier consistency is designed for complementary label learning,and the key problem of why MSE does not have classifier consistency under complementary label learning is explained.Experimental results show that the solutions proposed in this dissertation are comparable if not better than current state-of-the-art methods.