Friday PM, 28th August 2020
Gabriela Csurka is a Principal Scientist at NAVER LABS Europe, France. Her main research interests are in computer vision for image understanding, multi-view 3D reconstruction, visual localization, multi-modal information retrieval as well as domain adaptation (DA) and transfer learning. She has contributed to around 100 scientific communications, several on the topic of DA, received the best paper award at the Transferring and Adapting Source Knowledge in Computer Vision Workshop (TaskCV) in 2016, participated with success in DA related challenges (ImageClefDA’14, VisDA’17), and has given invited talks on domain adaptation (ACIVS’15, Task-CV’17, OpenMIC’18 and Task-CV’19). In 2017 she edited a book on Domain Adaptation in Computer Vision Applications.
Timothy M. Hospedales is a Reader at University of Edinburgh; Principal Scientist at Samsung AI Research Centre, Cambridge; and Alan Turing Institute Fellow. His research focuses on lifelong machine learning, broadly defined to include multi-domain/multi-task learning, domain adaptation, transfer learning and meta-learning; with applications including computer vision, vision and language, reinforcement learning for control and finance. He has co-authored numerous papers on domain adaptation, domain generalisation, and transfer learning in major venues including CVPR, ICCV, ECCV, ICML and AAAI. He teaches computer vision at Edinburgh University and has given invited talks on these topics at Task-CV and Deep Learning in Finance Summit as well as tutorials at ACM Multimedia and several Summer Schools.
Mathieu Salzmann is a Senior Researcher at EPFL, with a broad expertise in Computer Vision and Machine Learning. He has published several articles at major conferences and journals on the topic of Domain Adaptation, and has contributed a chapter on the topic of matching distributions in G. Csurka’s book on Domain Adaptation in Computer Vision Applications. Furthermore, he has been invited to present his domain adaptation work at various venues, including the Workshop on Domain Adaptation and Few-Shot Learning, the University of Oxford and ETHZ.
Tatiana Tommasi is an Assistant Professor at Politecnico di Torino, Italy and an affiliated researcher at the Italian Institute of Technology. She pioneered the area of transfer learning for computer vision and has large experience in domain adaptation, generalization and multimodal learning with applications for robotics and medical imaging. Tatiana received the best paper award at the 1st edition of Task-CV workshop at ECCV’14 and since then she has been leading the organization of the following workshop editions. She also organized a workshop on similar topics at NIPS’13,’14 and taught a tutorial at ECCV’14.
While huge volumes of unlabeled data are generated and made available in many domains, the cost of acquiring data labels remains high. On the other hand, solving problems with deep neural networks has become extremely popular, however current methods typically rely on massive amounts of labeled training data to achieve high performance. To overcome the burden of annotation, solutions have been proposed in the literature to exploit available unlabeled data from the same domain, referred to as semi-supervised learning; and to exploit labeled data or trained models available in similar, yet different domains, referred to as domain adaptation. The focus of this tutorial will be on the latter. Domain adaptation is also of increasing societal importance as vision systems are deployed in mission critical applications whose predictions have real-world impact, but where real-world testing data statistics can differ significantly from lab collected training data. Our aim will be to give an overview of visual domain adaptation methods, a field whose popularity in the computer vision community has increased significantly in the last few years, as attested by the proliferation of DA-related papers published during the last years in top-ranked computer vision and machine learning conferences.
1. In the first part, we will present domain adaptation from a theoretical point of view. In particular, we will define the domain shift problem and illustrate its importance in computer vision. We will then introduce different ways to measure the distribution mismatch between two domains, referred to as the source and target domains. Specifically, we will review different distance metrics between probability distributions, such as the Maximum Mean Discrepancy and the Hellinger distance, as well as different ways to represent the source and target data, for instance using subspaces, so as to compare them. We will further explain how these techniques have been used in the past to design domain adaptation algorithms. In this context, we will first review the main historical contributions in domain adaptation, and then briefly study how these contributions have been translated to deep networks.
2. In the second part of the tutorial we will first discuss and compare different domain adaptation strategies that exploit deep architectures for visual recognition, such shallow models used with pretrained or fine-tuned deep features and deep architectures designed for domain adaptation. Then, we will overview recent trends in domain adaptation, including deep discriminative models with various discrepancy based and adversarial based losses, generative 2 and encoder-decoder based models, network parameter adaptation methods, semi-supervised and curriculum learning based models. We will present methods proposed in the literature for image classification, semantic segmentation, object detection and others.
3. In the third part, we will discuss all those particular cases that differ from the standard domain adaptation setting: source and target data may not cover exactly the same set of classes (partial, open-set), the target data may come as an online stream rather than being available altogether (continuous), several sources may be provided with different annotations levels (deep-cocktail, multi-source, predictive domain adaptation). Finally, domain generalization is the most challenging condition where no target data is available at training time. We will also relate domain adaptation and generalization with self-supervised learning.
4. The tutorial will conclude with an ending part dedicated to unifying perspectives and outlook. We will present deep tensor methods and meta-learning methods that provide frameworks to link domain adaptation and domain generalisation with related research topics including multi-task/multi-domain learning and few-shot learning. We will draw connections to related issues such as adversarial robustness, and further applications such as SBIR, VQA and deep RL.