Note that this workshop is a part of Asian Conference on Machine Learning 2021, and thus in order to attend the workshop, you should first register for the conference. Then, the link of the zoom meeting room can be found at the virtual site and the gather town of ACML 2021.
Program
Schedule
The workshop will use for scheduling, and it will be combined with invited talks and contributed talks.
| Time | Event |
|---|---|
| Opening Ceremony | |
| Host: Masashi Sugiyama | |
| Invited Talk 1 | |
| Title: Self-Supervised Learning Disentangled Group Representation as Feature | |
| Speaker: Hanwang Zhang | |
| Contributed talks | |
| Title: Is Complementary Label Learning a Very Difficult Task? Speaker: Deng-Bao Wang | |
| Title: Graph Poisson Networks: Semi-Supervised Learning with Extremely Limited Labels Speaker: Sheng Wan | |
| Title: Learning with Noisy Labels Revisited: A Study Using Real-world Human Annotations Speaker: Jiaheng Wei | |
| Title: Taming Overconfident Predictions on Unlabelled data from Hindsight Speaker: Jing Li | |
| Invited Talk 2 | |
| Title: Label-Efficient Learning of Vision Transformer | |
| Speaker: Boqing Gong | |
| Invited Talk 3 | |
| Title: Leveraged Weighted Loss for Partial Label Learning | |
| Speaker: Zhouchen Lin |
Invited Speakers
Zhouchen Lin, Peking University, China
Hanwang Zhang, Nanyang Technological University, Singapore
Boqing Gong, Research Scientist Google, US
Topics
Overview
Machine learning should not be accessible only to those who can pay. Specifically, modern machine learning is migrating to the era of complex models (e.g., deep neural networks), which require a plethora of well-annotated data. Giant companies have enough money to collect well-annotated data. However, for startups or non-profit organizations, such data is barely acquirable due to the cost of labeling data or the intrinsic scarcity in the given domain. These practical issues motivate us to research and pay attention to weakly supervised learning (WSL), since WSL does not require such a huge amount of annotated data. We define WSL as the collection of machine learning problem settings and algorithms that share the same goals as supervised learning but can only access to less supervised information than supervised learning. In this workshop, we discuss both theoretical and applied aspects of WSL.
This workshop is a series of our previous workshops at ACML 2019, SDM 2020, ACML 2020, and IJCAI 2021. Our particular emphasis at this workshop is incomplete supervision, inexact supervision, inaccurate supervision, cross-domain supervision, imperfect demonstration, and weak adversarial supervision (new topic).
Topics of Interest
WSL workshop includes (but not limited to) the following topics:
- Algorithms and theories of incomplete supervision, e.g., semi-supervised learning, active learning and positive-unlabeled learning;
- Algorithms and theories of inexact supervision, e.g., multi-instance learning and complementary learning;
- Algorithms and theories of inaccurate supervision, e.g., crowdsourced learning and label-noise learning;
- Algorithms and theories of cross-domain supervision, e.g., zero-/one-/few-shot learning, transferable learning and multi-task leaning;
- Algorithms and theories of imperfect demonstration, e.g., inverse reinforcement learning and imitation learning with non-expert demonstrations;
- Algorithms and theories of adversarial weakly-supervised learning, e.g., adversarial semi-supervised learning and adversarial contrastive learning and adversarial label-noisy learning;
- Broad applications of weakly supervised learning, such as weakly supervised object detection, weakly supervised sequence modeling, weakly supervised cross-media retrieval, and weakly supervised medical image segmentation.
Further Descriptions
The focus of this workshop is six types of weak supervision: incomplete supervision, inexact supervision, inaccurate supervision, cross-domain supervision, imperfect demonstration, and weak adversarial supervision. Specifically, incomplete supervision considers a subset of training data given with ground-truth labels while the other data remain unlabeled, such as semi-supervised learning and positive-unlabeled learning. Inexact supervision considers the situation where some supervision information is given but not as exacted as desired, i.e., only coarse-grained labels are available. For example, if we are considering to classify every pixel of an image, rather than the image itself, then ImageNet becomes a benchmark with inexact supervision. Besides, multi-instance learning belongs to inexact supervision, where we do not exactly know which instance in the bag corresponds to the given ground-truth label. Inaccurate supervision considers the situation where the supervision information is not always the ground-truth, such as label-noise learning.
Cross-domain supervision considers the situation where the supervision information is scarce or even non-existent in the current domain but can be possibly derived from other domains. Examples of cross-domain supervision appear in zero-/one-/few-shot learning, where external knowledge from other domains is usually used to overcome the problem of too few or even no supervision in the original domain. Imperfect demonstration considers the situation for inverse reinforcement learning and imitation learning, where the agent learns with imperfect or non-expert demonstrations. For example, AlphaGo learns a policy from a sequence of states and actions (expert demonstration). Even if an expert player wins a game, it is not guaranteed that every action in the sequence is optimal.
Weak adversarial supervision considers the situation where weak supervision meets adversarial robustness. Since ML models are increasingly deployed in real-world applications, AI security attracts more and more attention from both academia and industry. Therefore, many robust learning algorithms aim to prevent various evasion attacks, e.g., adversarial attacks, privacy attacks, model stealing attacks, and so on. However, almost all those robust algorithms (against evasion attacks) implicitly assume the strong supervision signals, which hardly meets the requirements in practice. For example, almost all adversarial training algorithms assume the labels of training data are clean without any noisy labels, which is not true in practice even in the example of benchmark datasets like CIFAR-10 and MNIST. Therefore, when we develop evasion-robust algorithms, it is very practical/urgent to consider the supervision signals are imperfect.
This workshop will discuss the fundamental theory of weakly supervised learning. Although theories of weakly supervised statistical learning already exist, extending these results for weakly supervised learning is still a challenge. Besides, this workshop also discusses on broad applications of weakly supervised learning, such as weakly supervised object detection (computer vision), weakly supervised sequence modeling (natural language processing), weakly supervised cross-media retrieval (information retrieval), and weakly supervised medical image segmentation (healthcare analysis).
Organizers
Program Co-chairs
Jingfeng Zhang, RIKEN, Japan.
Feng Liu, The University of Technology Sydney, Australia.
Nan Lu, The University of Tokyo, Japan.
Lei Feng, Chongqing University, China.
Tongliang Liu, The University of Sydney, Australia.
Bo Han, Hong Kong Baptist University, Hong Kong SAR, China.
Gang Niu, RIKEN, Japan.
Masashi Sugiyama, RIKEN / University of Tokyo, Japan.
Advisory Board (alphabetical order by last name)
Chen Gong, Nanjing University of Science and Technology, China.
Mingming Gong, The University of Melbourne, Australia.
Yang Liu, University of California, Santa Cruz, US.
Ivor W. Tsang, University of Technology Sydney, Australia.
Yisen Wang, Peking University, China.
Miao Xu, The University of Queensland, Australia.
Quanming Yao, Tsinghua University / 4Paradigm Inc., China.
Min-Ling Zhang, Southeast University, China.
Previous Workshops
IJCAI2021 WSRL Workshop, Online.
ACML2020 WSRL Workshop, Online.
SDM2020 WSUL Workshop, Ohio, United States.
ACML2019 WSL Workshop, Nagoya, Japan.