Open AccessPosted Content
A Review of Single-Source Deep Unsupervised Visual Domain Adaptation
Sicheng Zhao,Xiangyu Yue,Shanghang Zhang,Bo Li,Han Zhao,Bichen Wu,Ravi Krishna,Joseph E. Gonzalez,Alberto Sangiovanni-Vincentelli,Sanjit A. Seshia,Kurt Keutzer +10 more
TL;DR: This article review the latest single-source deep unsupervised DA methods focused on visual tasks and discusses new perspectives for future research, including discrepancy-based methods, adversarial discriminative methods, and self-supervision- based methods.
read more
Abstract: Large-scale labeled training datasets have enabled deep neural networks to excel across a wide range of benchmark vision tasks. However, in many applications, it is prohibitively expensive and time-consuming to obtain large quantities of labeled data. To cope with limited labeled training data, many have attempted to directly apply models trained on a large-scale labeled source domain to another sparsely labeled or unlabeled target domain. Unfortunately, direct transfer across domains often performs poorly due to the presence of domain shift or dataset bias. Domain adaptation is a machine learning paradigm that aims to learn a model from a source domain that can perform well on a different (but related) target domain. In this paper, we review the latest single-source deep unsupervised domain adaptation methods focused on visual tasks and discuss new perspectives for future research. We begin with the definitions of different domain adaptation strategies and the descriptions of existing benchmark datasets. We then summarize and compare different categories of single-source unsupervised domain adaptation methods, including discrepancy-based methods, adversarial discriminative methods, adversarial generative methods, and self-supervision-based methods. Finally, we discuss future research directions with challenges and possible solutions.
read more
Chat with Paper
AI Agents for this Paper
Find similar papers on Google Scholar, PubMed and Arxiv
Write a critical review of this paper
Analyze citations of this paper to find unaddressed research gaps
Figures
Fig. 1. An example of domain shift. For both image-level object classification and pixel-wise semantic segmentation tasks, direct transfer of the models trained on the labeled source domain to the unlabeled target domain results in a dramatic performance drop. 
TABLE III COMPARISON OF DIFFERENT DISCREPANCY-BASED METHODS, WHERE ‘DISCREPANCY’ INDICATES THE DISCREPANCY LOSS, ‘LOSS LEVEL’ INDICATES WHAT LEVEL THE LOSS IS APPLIED TO, ‘LAYER’ REPRESENTS THE LAYERS THAT THE LOSS FUNCTIONS ON, ‘WEIGHT’ INDICATES WHETHER THE WEIGHTS OF THE TWO NETWORKS ARE SHARED OR NOT, ‘DISTRIBUTION’ INDICATES WHAT TYPE OF DISTRIBUTION IS ALIGNED. 
TABLE VII COMPARISON OF DIFFERENT SINGLE-SOURCE DUDA CATEGORIES. (THE MORE STARS THE METHOD HAS, THE BETTER IT IS. ) 
TABLE VIII PERFORMANCE COMPARISON (CLASSIFICATION ACCURACY IN %) OF DIFFERENT METHODS ON DIGIT DATASET FOR DIGIT RECOGNITION. ‘BACKBONE’ DENOTES THE BASE NETWORK ARCHITECTURE, ‘M’, ‘M-M’, ‘U’, ‘S’ ARE DIFFERENT DOMAINS (SEE SECTION III FOR DETAILS), AND ‘–>’ REPRESENTS THE ADAPTATION FROM ONE SOURCE DOMAIN TO ANOTHER TARGET DOMAIN. THE COLUMN ‘C’ INDICATES WHICH CATEGORY THE METHOD BELONGS TO, WHERE ‘D’, ‘A’, ‘G’, ‘S’, ‘O’ ARE RESPECTIVELY SHORT FOR DISCREPANCY-BASED, ADVERSARIAL DISCRIMINATIVE, ADVERSARIAL GENERATIVE, SELF-SUPERVISION-BASED METHODS, AND OTHERS (THE SAME BELOW). 
Fig. 2. Classification of widely employed framework of different single-source deep unsupervised domain adaptation (DUDA) pipelines. Most existing methods can be obtained by employing different component values, slightly changing the architecture, or combining different pipelines. 
TABLE IV COMPARISON OF DIFFERENT ADVERSARIAL DISCRIMINATIVE MODELS, WHERE ‘EN’ IS SHORT FOR ENCODER. ADVERSARIAL LEVEL REFERS TO THE LEVEL OF ALIGNMENT FOR THE DISCRIMINATOR INPUT, EITHER GLOBALLY OR CLASS-WISELY.
Citations
Transferable Semantic Augmentation for Domain Adaptation
Shuang Li,Mixue Xie,Kaixiong Gong,Chi Harold Liu,Yulin Wang,Wei Li +5 more
- 01 Jun 2021
TL;DR: Transferable Semantic Augmentation (TSA) as mentioned in this paper is proposed to enhance the classifier adaptation ability through implicitly generating source features towards target semantics, where the source features can be augmented to effectively equip with target semantics to train a more transferable classifier.
Deep Transfer Learning for Bearing Fault Diagnosis: A Systematic Review Since 2016
TL;DR: In this paper , a comprehensive review of the development of deep transfer learning-based bearing fault diagnosis approaches since 2016 is presented, which provides a systematic guideline for researchers and practitioners to efficiently identify suitable DTL models based on the actual problems encountered in bearing fault prediction.
125
Emotion Recognition From Multiple Modalities: Fundamentals and methodologies
TL;DR: In this article, the authors discuss several key aspects of multimodal emotion recognition (MER) and present a tutorial on how to recognize, interpret, process, and simulate emotions.
107
Transfer learning in environmental remote sensing
Yuchi Ma,Shuo Chen,Stefano Ermon,David B. Lobell +3 more
TL;DR: This systematic review of 1676 papers (2017-2022) explores transfer learning in environmental remote sensing, highlighting successes and gaps in applications like land cover mapping, vegetation monitoring, and natural disaster management, and identifies research directions for improvement.
106
•Posted Content
ePointDA: An End-to-End Simulation-to-Real Domain Adaptation Framework for LiDAR Point Cloud Segmentation
TL;DR: ePointDA is proposed, a novel end-to-end framework that enables ePointDA to bridge the domain shift at the pixel-level by explicitly rendering dropout noise for synthetic LiDAR and at the feature- level by spatially aligning the features between different domains, without requiring the real-world statistics.
103
References
Graph Matching and Pseudo-Label Guided Deep Unsupervised Domain Adaptation
Debasmit Das,C. S. George Lee +1 more
- 04 Oct 2018
TL;DR: This paper proposes a novel graph-matching metric between the source and target domain representations and proposes a solution to unsupervised domain adaptation where the source domain data is labelled but unlabelled target domain data.
EmotionGAN: Unsupervised Domain Adaptation for Learning Discrete Probability Distributions of Image Emotions
Sicheng Zhao,Xin Zhao,Guiguang Ding,Kurt Keutzer +3 more
- 15 Oct 2018
TL;DR: A novel adversarial model for emotion distribution learning is developed, termed EmotionGAN, which alternately optimizes the Generative Adversarial Network (GAN) loss, semantic consistency loss, and regression loss and can adapt source domain images such that they appear as if they were drawn from the target domain, while preserving the annotation information.
•Proceedings Article
Multiple source adaptation and the Rényi divergence
Yishay Mansour,Mehryar Mohri,Afshin Rostamizadeh +2 more
- 18 Jun 2009
TL;DR: This paper presents a novel theoretical study of the general problem of multiple source adaptation using the notion of Renyi divergence, extending previous multiple source loss guarantees based on distribution weighted combinations to arbitrary target distributions P, not necessarily mixtures of the source distributions, and proving a lower bound.
•Proceedings Article
Domain adaptation with conditional distribution matching and generalized label shift
Remi Tachet des Combes,Han Zhao,Yu-Xiang Wang,Geoffrey J. Gordon +3 more
- 10 Dec 2020
TL;DR: This article proposed generalized label shift (GLS) as a way to improve robustness against mismatched label distributions, which improves the transfer performance of adversarial domain adaptation methods in multi-class classification and general discriminators.
•Proceedings Article
Towards Fast Computation of Certified Robustness for ReLU Networks
Tsui-Wei Weng,Huan Zhang,Hongge Chen,Zhao Song,Cho-Jui Hsieh,Duane S. Boning,Inderjit S. Dhillon,Luca Daniel +7 more
- 25 Apr 2018
TL;DR: In this paper, the authors exploit the special structure of ReLU networks and provide two computationally efficient algorithms FastLin and Fast-Lip that are able to certify non-trivial lower bounds of minimum distortions, by bounding the ReLU units with appropriate linear functions Fast-Linear or FastLip.