Structurally Regularized Unsupervised Learning And Domain Adaptation

In the age of big data,one can use the knowledge from existing data related to the target task to help model learning for the target task with no labeled data,i.e.,transfer learning.Denote the domain of previous data as the source domain,and the domain of current task as the target domain.When the source and target domains have different distributions but share the label space,the transfer learning problem is equivalent to the domain adaptation problem.This dissertation studies unsupervised domain adaptation(UDA),where target samples are all unlabeled.In essence,UDA provides geometric structural constraints from task characteristics or decision space to help learn decision boundaries of unlabeled data,where geometric structural constraints are provided by labeled source domain data.On the other hand,now it is becoming more and more important to work on methods that use simulated data but perform well in practical domains whose data or annotation are difficult to acquire,e.g.,self-driving,medical imaging,and 3D vision.Therefore,this dissertation also specifically studies the typical scenario of domain adaptation,i.e.,synthetic-to-real transfer.This dissertation first introduces the background and significance of the research,reviews the current research context,and points out advantages and limitations of various methods,particularly the mainstream methods for UDA—non-generative model based adversarial methods.Then,this dissertation focuses on unsupervised learning settings,and mainly does methodological and empirical studies on transfer learning,domain adaptation,and synthetic-to-real transfer for vision semantic tasks such as image classification and semantic segmentation;by proposing new methods to improve discriminative structures of data,this dissertation can significantly improve the recognition accuracy.The details are as follows:(1)To solve the problem of mode collapse caused by the defective design of network structure and loss function and limited information interaction in mainstream methods,this dissertation proposes Discriminative Adversarial Domain Adaptation(DADA),which improves the alignment of joint distributions of feature and category across domains.It alleviates the mode collapse problem and significantly enhances the recognition accuracy.Experimental results verify its efficacy and superiority,e.g.,under partial adaptation setting on Office-Home dataset,the recognition accuracy is improved by 10.6%.(2)To solve the problem that mainstream methods do not fully depict intrinsic discriminative structures of target data and thus have a limited ability in category-level domain alignment,this dissertation proposes a new method of Vicinal and Categorical Domain Adaptation(Vi Cat DA)and designs novel adversarial losses at multiple levels on both source and target domains.It enhances the ability of category-level domain alignment and improves the classification performance on target test samples.Experimental results show that Vi Cat DA achieves significantly better performance than existing methods,e.g.,the recognition accuracy is improved by 2.2% on Vis DA-2017 dataset.(3)To address the risk of damaging intrinsic data structures of target discrimination and the problem of insufficient generalization in mainstream methods,this dissertation proposes a hybrid model of Structurally Regularized Deep discriminative Clustering(H-SRDC).It integrates discriminative clustering with a generative one,regularized by the source data structure.It can uncover the intrinsic target discrimination and greatly improve the generalization to out-of-distribution test samples.A comprehensive evaluation of different UDA methods in both inductive and transductive settings finds that H-SRDC consistently outperforms existing methods in both settings,e.g.,the recognition accuracy is improved by 5.4% in the inductive setting on Office-Home dataset.(4)To solve the problem that mainstream methods do not consider the basic assumption of domain adaptability and do not learn a pure,compact feature representation for each class,this dissertation is motivated by the basic assumption of domain adaptability and proposes a new method of distilled discriminative clustering for domain adaptation(termed Dis Cluster DA).This method helps learn class-wisely pure,compact feature distributions and achieves considerable performance gain.Experiments verify the efficacy and superiority of Dis Cluster DA,e.g.,the classification accuracy is improved by 6.7% on Vis DA-2017 dataset.(5)To solve the basic and important problems in the context of image classification,e.g.,lack of comprehensive synthetic data research and insufficient exploration of synthetic-to-real transfer,this dissertation utilizes 3D rendering and domain randomization to generate synthetic datasets and do a comprehensive study on both supervised learning and downstream transferring.Synthetic data is used to explore the model generalization problem in image classification and a more large-scale synthetic-to-real benchmark is introduced for classification adaptation(termed S2RDA),which can push forward the future DA research.