A Study On Open-Set Deep Semi-Supervised Learning

Semi-Supervised Learning(SSL),as an important branch of weakly supervised learning,aims to solve the realistic situation when most of the label information is missing.In practice,the acquirement of labels often consumes unaffordable human labor,monetary costs as well as material resources.Therefore,the obtained labels are often extremely limited.SSL can effectively utilize scarce labeled data,meanwhile taking the most advantage of abundant unlabeled data,so as to train an accurate classifier.Nonetheless,traditional SSL relies on the closed-set assumption,which assumes that the labeled data and unlabeled data are drawn from the exact same distribution.However,in the open-set situation,the distribution of labeled data and unlabeled data often differs in both class and feature.Such difference would seriously hinder the performance of traditional SSL.Hence,for better practical application of SSL under open-set situations,this dissertation aims to make contributions regarding the following two aspects:1)First,this dissertation considers Subset Class Mismatch problem,which is a specific open-set case denoting the classes in labeled data are only a subset of unlabeled data,and the unlabeled data contain some classes that are private from labeled data.To solve this problem,this dissertation proposes Transferable OOD data Recycling(TOOR),which can detect the unlabeled private classes,in order to largely exploit the information contained by unlabeled data,meanwhile progressively recycle the unlabeled data which can contribute to the network learning.Extensive experiments have shown that TOOR can perfectly tackle the subset class mismatch problem and achieve excellent SSL performance.2)Because focusing on a certain open-set case still has some distance to the real application,therefore,this dissertation proposes to consider different situations of the open-set problem,which includes Class Mismatch and Feature Mismatch.For class mismatch,the investigated situation involves not only the aforementioned subset class mismatch,but also Intersectional Class Mismatch,which denotes the classes in labeled data and those in unlabeled data are intersectional.Moreover,the feature mismatch denotes that the labeled data and unlabeled data are sampled from different feature distributions.To solve such problem,this dissertation proposes Class-sh Aring data detection and Feature Adaptation(CAFA),which utilizes a novel evaluation strategy to weigh each datum so as to detect the class-sharing data,thus it can universally deal with different cases in open-set problem.Furthermore,CAFA conducts adversarial training to mitigate the distribution difference between labeled data and unlabeled data,which guarantees that SSL can fully exploit the potential value of unlabeled data.Extensive experiments show that CAFA can be applied to various SSL backbone methods and achieve encouraging results.