The 1st Visual Intelligence Seminar on Few-shot Learning

Yuxiong Wang is an Assistant Professor in the Department of Computer Science at the University of Illinois at Urbana-Champaign. He received a Ph.D. in robotics from Carnegie Mellon University under the supervision of Martial Hebert. His research interests lie in computer vision, machine learning, and robotics, with a particular focus on few-shot learning, meta-learning, and streaming perception. He has spent time at Facebook AI Research (FAIR). He received the Best Paper Honorable Mention Award in ECCV 2020.

Timothy Hospedales is a Full Professor of Artificial Intelligence at University of Edinburgh, Programme Director of Machine Learning and Principal Scientist at Samsung AI Research Centre, Cambridge, and Alan Turing Institute Fellow. His work has been funded by EU Horizon 2020, UK EPSRC, and industry and led to several best paper awards and nominations. He is associate editor of IEEE Transactions on Pattern Analysis and Machine Intelligence. His research interest is in meta-learning for data efficiency and robustness with applications in vision, language and control. He has published on these topics at major venues including CVPR, NeurIPS, ICML, AAAI, and authored a comprehensive survey on meta-learning.

In this talk I will first give an overview perspective and taxonomy of major work the field, as motivated by our recent survey paper on meta-learning in neural networks. I hope that this will be informative for newcomers, as well as reveal some interesting connections and contrasts that will be thought-provoking for experts. I will then give a brief overview of recent meta-learning work from my group which covers some broad issues in computer vision where meta-learning can be applied including dealing with domain-shift, data augmentation, learning with label noise, and accelerating RL. Finally, I will describe some recent results that may settle the recent debate in few-shot learning about whether or not meta-learning is made redundant by strong feature extractor training.

陈松灿，南京航天航空大学计算机科学和技术学院/人工智能学院教授。 国际模式识别学会会士 (IAPR Fellow)和中国人工智能学会会士(CAAI Fellow)。 Google Scholar被引数超14820次，H-指数55。2014-2019连续6年入选Elsevier中国高引学者榜。 现任中国人工智能学会机器学习专委会主任、常务理事和江苏省人工智能学会常务副理事长。 至今主持国家自然科学基金项目12项，其中重点项目1项。

自监督学习是一种渐受关注的无监督表示学习范式，近年来已在包括计算机视觉和自然语言处理等众多领域显示出应用的有效性。SSL通过利用数据自身的内在规则性等自动生成判别信息实现学习。本报告首先进行选择性回顾，其次对其在何条件下会对下游任务奏效进行初步探索，最后给出了展望。

Yi Zhu is an Applied Scientist in Amazon AI. Before that, he obtained the Ph.D. degree in Computer Science at UC Merced in 2019. His research interests include video understanding, semantic segmentation and self-supervised representation learning. He co-organized tutorials on reproducing SOTA deep learning models in ICCV 2019 and tutorials on video modeling in CVPR 2020. He is also an active contributor to open source projects, including GluonCV, MXNet and Nvidia-segmentation.

Zhiwu Lu is a full professor with the Gaoling School of Artificial Intelligence, Renmin University of China, Beijing 100872, China. He received the Master of Science degree in applied mathematics from Peking University in 2005, and the PhD degree in computer science from City University of Hong Kong in 2011. He have published over 70 papers in international journals and conference proceedings including TPAMI, IJCV, TIP, ICLR, NeurIPS, CVPR, ICCV, and ECCV. He won the IBM SUR Award 2015, and Best Paper Award at CGI 2014. His team took the 2nd place in the VID task of ILSVRC 2015.

Most recent few-shot learning (FSL) methods take a meta-learning framework. Since the performance of the SOTA FSL methods is saturating on benchmark datasets, new meta-learning paradigms are needed in this field. In this talk, I will introduce two such paradigms, MELR (Modeling Episode-Level Relationships) and IEPT (Instance-Level and Episode-Level Pretext Tasks), for standard inductive FSL setting. Specifically, MELR aims to explicitly model the episode-level relationships during meta-training, and IEPT aims to seamlessly integrate self-supervised learning (SSL) into supervised FSL. Both MELR and IEPT achieve new SOTA on several benchmarks. These two paradigms have been published as two ICLR 2021 papers.

Dr Li Zhang is a tenure-track Associate Professor at the School of Data Science, Fudan University. Previously, he was a Research Scientist at Samsung AI Center Cambridge, and a Postdoctoral Research Fellow at the University of Oxford. Prior to joining Oxford, he read his PhD in computer science at Queen Mary University of London.

Shangzhe Wu is a PhD student in the Visual Geometry Group at the University of Oxford, supervised by Andrea Vedaldi. His current research is focused on unsupervised 3D understanding. He received his bachelor’s degree in computer science from HKUST, where he worked with Chi-Keung Tang and Yu-Wing Tai on image translation. His work received the Best Paper Award at CVPR 2020.

Weidi is a research fellow at Visual Geometry Group, working on computer vision, biomedical image analysis.

Anurag Arnab is a research scientist at Google. Prior to this, he completed his PhD at the University of Oxford under the supervision of Professor Philip Torr. His research interests lie primarily in scene understanding problems in computer vision under various forms of supervision.

Prof. Tao Xiang a professor of Computer Vision and Machine Learning and University Distinguished Chair at the University of Surrey. He is also a Research Scientist at Facebook AI. He received his PhD degree from the National University of Singapore in 2002. Before joining University Surrey, he was a professor at Queen Mary University of London. Prof. Xiang’s research in computer vision has been focused on video surveillance, daily activity analysis, and sketch analysis. He also has interests in large-scale machine learning problems including zero/few-shot learning and domain adaptation. He has published over 200 papers with 19K citations (h-index 72). He served as area chairs for CVPR, ICCV and ECCV, and is an associated editor for IEEE Trans. on PAMI.

In this talk, I will first give a brief overview of the recent trend on deep learning with sparely labeled data and highlight several related research problems. These include few-shot learning, unsupervised domain adaptation and domain generalization. I will then give three examples of how these research problems can be tackled based on some of my recent works. These include: how an object detector can add a new class with as few as one example using a hyper-network; how stochastic neural networks can be utilized to solve the unsupervised domain adaptation problem; and how to exploit data synthesis to train more generalizable models.