20:00 -- 20:10 Welcome
20:10 -- 20:50
Keynote 1: Abnormal event detection and prediction across urban big data
Professor Yong Li, Tsinghua University, China
20:50 -- 21:30
Keynote 2: Object-centric anomaly detection in video
Professor Radu Tudor Ionescu, University of Bucharest, Romania
21:30 -- 21:40 Short Break
21:40 -- 22:40
Invited Talks from top conference papers (20 minutes per paper, including 5 minutes for Q&A):
Kaize Ding. Few-shot Network Anomaly Detection via Cross-network Meta-learning. In Proceedings of the Web Conference 2021.
Arizona State University, USA
2. Hyunsoo Cho. Masked Contrastive Learning for Anomaly Detection. In Proceedings of IJCAI 2021.
Seoul National University, South Korea
3. Yu Tian. Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning. In Proceedings of ICCV 2021.
University of Adelaide, Australia
10:00 -- 10:40
Keynote 3: A Perspective on Anomaly Detection
Professor Sanjay Chawla, Qatar Computing Research Institute, Qatar
10:40 -- 11:20
Keynote 4: On the difference between anomaly/novelty detection and formal open-set/open-world algorithms
Professor Terrance E. Boult, University of Colorado Colorado Springs, USA
11:20 -- 11:30 Short Break
11:30 -- 12:10
Keynote 5: Automating unsupervised outlier model selection
Professor Leman Akoglu, Carnegie Mellon University, USA
12:10 -- 12:50
Keynote 6: Uncovering the Unknowns of Deep Neural Networks: Challenges and Opportunities
Professor Sharon Yixuan Li, University of Wisconsin-Madison, USA
12:50 -- 13:00 Short Break
13:00 -- 14:00
Presentation of Accepted Papers (20 minutes per paper, including 5 minutes for Q&A):
Sara Hajj Ibrahim and Mohamed Nassar. Hack The Box: Fooling Deep Learning Abstraction-Based Monitors.
American University of Beirut (Lebanon), University of New Haven (USA)
2. Konstantin Kirchheim, Tim Gonschorek and Frank Ortmeier. Addressing Randomness in Evaluation Protocols for Out-of-Distribution Detection.
Otto von Guericke University Magdeburg, Germany
Keynote 1: Abnormal event detection and prediction across urban big data
Abstract: Urban abnormal events such as traffic incidents and unexpected crowds pose a significant threat to social order and public safety. Alerting abnormal events in their early stages or even predicting the happening of such events are of great value for emergency handling and anomaly control. In recent years, with the fast development of mobile smart devices and ubiquitous sensing techniques, spatio-temporal big data are continuously produced in cities. Encouraged by the wide access to urban big data, AI techniques for urban anomalous event detection and prediction have become a new and unique topic in the anomaly detection research area. In this talk, we first quickly review the concepts and challenges of the urban anomaly analysis problem. Then, we introduce two of our works on urban abnormal event detection and prediction. In the first work, we proposed an urban data decomposition method to extract abnormal influence of public emergencies for event detection. In the second work, we predict the occurring of urban anomalies considering the impacts of past events and urban environmental factors. Especially, we model the impact of historical events as the message passing process among urban regions. Finally, we discuss potential research directions from aspects of data challenges and applications.
Speaker: Yong Li, ,Tsinghua University
Bio: Dr. Li is currently a Tenured Associate Professor of the Department of Electronic Engineering, Tsinghua University. He received the Ph.D. degree in electronic engineering from Tsinghua University in 2012. His research interests include machine learning and big data mining, particularly, automatic machine learning and spatial-temporal data mining for urban computing, recommender systems, and knowledge graphs. Dr. Li has served as General Chair, TPC Chair, SPC/TPC Member for several international workshops and conferences, and he is on the editorial board of two IEEE journals. He has published over 100 papers on first-tier international conferences and journals, including KDD, NeurIPS，WWW, UbiComp, SIGIR, AAAI, TKDE, TMC etc, and his papers have total citations more than 12000. Among them, ten are ESI Highly Cited Papers in Computer Science, and five receive conference Best Paper (run-up) Awards. He received IEEE 2016 ComSoc Asia-Pacific Outstanding Young Researchers, Young Talent Program of China Association for Science and Technology, and the National Youth Talent Support Program.
Keynote 2: Object-centric anomaly detection in video
Abstract: Anomaly detection in video is a challenging computer vision problem, as the classification of an event as normal or abnormal always depends on context. For instance, driving a truck on the street is considered normal, but, if the truck enters a pedestrian area, the event becomes abnormal. Considering the commonly adopted definition of abnormal events and the reliance on context, it is difficult to obtain a sufficiently representative set of anomalies for all possible contexts, making traditional supervised methods less applicable to abnormal event detection. In this talk, we will present a series of recent anomaly detection methods that are trained without direct supervision. The presented methods propose alternative object-centric approaches such as designing proxy self-supervised tasks or using pseudo-abnormal examples in an adversarial fashion. By focusing strictly on objects, object-centric approaches can significantly reduce false detection rates, forming a viable solution for real world scenarios.
Speaker: Radu Tudor Ionescu, University of Bucharest
Bio: Radu Tudor Ionescu is Professor at the University of Bucharest, Romania. He completed his PhD at the University of Bucharest in 2013. He received the 2014 Award for Outstanding Doctoral Research in the field of Computer Science from the Romanian Ad Astra Association. His research interests include machine learning, computer vision, image processing, medical imaging, text mining and computational linguistics. He published over 90 articles at international peer-reviewed conferences (e.g.: CVPR, NeurIPS, ICCV, ACL, EMNLP, NAACL, EACL, ECML-PKDD, WACV, INTERSPEECH) and journals (e.g.: IEEE TPAMI, COLI, CVIU, NN), and a research monograph with Springer. Radu received the "Caianiello Best Young Paper Award" at ICIAP 2013 for the paper entitled "Kernels for Visual Words Histograms". He also received the "Young Researchers in Science and Engineering" Prize and the "Danubius Young Scientist Award 2018 for Romania" from the Austrian Federal Ministry of Education, Science and Research and by the Institute for the Danube Region and Central Europe. Together with other co-authors, he participated at several international competitions. They have ranked on 4th place in the Facial Expression Recognition Challenge of the WREPL Workshop of ICML 2013, 3rd place in the Native Language Identification Shared Task of the BEA-8 Workshop of NAACL 2013, 2nd place in the Arabic Dialect Identification Shared Task of the VarDial Workshop of COLING 2016, 1st place in the Arabic Dialect Identification Shared Task of the VarDial Workshop of EACL 2017, 1st place in the Native Language Identification Shared Task of the BEA-12 Workshop of EMNLP 2017, 1st place in the Arabic Dialect Identification Shared Task of the VarDial Workshop of COLING 2018.
Keynote 3: A Perspective on Anomaly Detection
Abstract: Using a few use cases and a series of questions that are typically asked about anomaly detection, the talk will provide a personal perspective on the research landscape in anomaly detection.
Speaker: Sanjay Chawla, Qatar Computing Research Institute
Bio: Sanjay Chawla is a research director at the Qatar Computing Research Institute (QCRI). Prior to joining QCRI, Dr. Chawla was a Professor in the Faculty of Engineering and IT at the University of Sydney. From 2008-2011, he also served as the Head (Department Chair) of the School of Information Technologies. While his work spans many areas of data science, research in anomaly detection has been a consistent theme. Dr. Chawla's work has been recognized with several best paper awards, including in leading conferences such as SIAM International Conference in Data Mining (2006) and IEEE International Conference in Data Mining (2010). Most recently, and along with his ex-PhD student and now Professor Wei Liu, his work received the "most influential paper award" at PAKDD 2021. He served as PC Chair of KDD (2021) and PAKDD (2012).
Keynote 4: On the difference between anomaly/novelty detection and formal open-set/open-world algorithms
Abstract: The first part of the talk will start with Dr. Boult's Motivation for studying Anomaly and then formal Open-Set algorithms and dealing with "unknowns" inputs in fielded systems. We then summarize our formalization of Open-Set and Open world problems and how that differs from simple rejection techniques for out-of-distribution detection (OOD). The second part of the talk presents a range of our algorithms which are formally open-set algorithms and their key properties. These include variations on classic algorithms, W-SVM, PI-SVM, and most recently, Specialized SVM. We also summarize our Open World algorithms NNO and Extreme-value machine, highlighting the latter with our recent work on open-world learning without labels. Finally, the talk addresses issues with evaluation metrics for these classes of problems.
Speaker: Terrance E. Boult, University of Colorado Colorado Springs, USA
Bio: Dr. Terry Boult, Distinguished El Pomar Professor of Innovation and Security at University of Colorado Colorado Springs (UCCS), researches computer vision, machine learning, biometrics, and security. Before joining UCCS in 2003, he was an endowed professor and founding chairman of Lehigh University's CSE Department and from 1986-1992 was a faculty at Columbia University. Dr. Boult is an IEEE fellow with hundreds of papers, 15 patents, has been involved in multiple successful start-up companies in the security space, and developed software/systems deployed worldwide. He is an innovator with a passion for combining teaching, research, and business and has won multiple teaching, research, innovation, and entrepreneurial awards. At the University of Colorado at Colorado Springs, he was the architect for the awarding-winning Bachelor of Innovation(TM) family of degrees and a key member in founding the UCCS Ph.d. in Engineering Security. More details can be found at www.vast.uccs.edu/~tboult/vita.html
Keynote 5: Automating unsupervised outlier model selection
Abstract: Given an unsupervised outlier detection task on a new dataset, how can we automatically select a good outlier detection algorithm and its hyperparameter(s) (collectively called a model)? In this talk, I will present METAOD, a principled, data-driven approach to unsupervised outlier model selection (UOMS) problem based on meta-learning. UOMS is notoriously challenging, as compared to model selection for classification and clustering, since (i) model evaluation is infeasible due to the lack of hold-out data with labels, and (ii) model comparison is infeasible due to the lack of a universal objective function. METAOD capitalizes on the performances of a large body of detection models on historical outlier detection benchmark datasets, and carries over this prior experience to automatically selecting an effective model to be employed on a new dataset without any labels, model evaluations or model comparisons. To capture task similarity within our meta-learning framework, we introduce specialized meta-features that quantify outlying characteristics of a dataset. Extensive experiments show that selecting a model by METAOD significantly outperforms no model selection (e.g. always using the same popular model or the ensemble of many) as well as other meta-learning techniques tailored for UOMS. Moreover, upon (meta-)training, METAOD is extremely efficient at test time; selecting from a large pool of 300+ models in less than 1 second for a new task. METAOD and our meta-learning database is open-sourced for practical use and to foster further research on the UOMS problem, at http://bit.ly/MetaOD
Speaker: Leman Akoglu, Carnegie Mellon University
Bio: Leman Akoglu is the Heinz College Dean's Associate Professor of Information Systems, Carnegie Mellon University. She also holds courtesy appointments in the Computer Science Department (CSD) and the Machine Learning Department (MLD) of School of Computer Science (SCS). Prior to this, she was an Assistant Professor in the Department of Computer Science at Stony Brook University since receiving her Ph.D. from CSD/SCS of Carnegie Mellon University in 2012. Dr. Akoglu’s research interests broadly span machine learning and graph mining, focusing on pattern discovery and anomaly mining, with applications to fraud and event detection. At Heinz, Dr. Akoglu directs the Data Analytics Techniques Algorithms (DATA) Lab. Dr. Akoglu is a recipient of the SDM/IBM Early Career Data Mining Research award (2020), National Science Foundation CAREER award (2015) and US Army Research Office Young Investigator award (2013). Her research has won 7 publication awards; Best Research Paper at SIAM SDM 2019, Best Student Machine Learning Paper Runner-up at ECML PKDD 2018, Best Paper Runner-up at SIAM SDM 2016, Best Research Paper at SIAM SDM 2015, Best Paper at ADC 2014, Best Paper at PAKDD 2010, and Best Knowledge Discovery Paper at ECML PKDD 2009. She also holds 3 U.S. patents filed by IBM T. J. Watson Research Labs. Her research has been supported by the NSF, US ARO, DARPA, Adobe, Facebook, Northrop Grumman, PNC Bank, and PwC.
Keynote 6: Uncovering the Unknowns of Deep Neural Networks: Challenges and Opportunities
Abstract: The real world is open and full of unknowns, presenting significant challenges for machine learning (ML) systems that must reliably handle diverse, and sometimes unknown inputs. Out-of-distribution (OOD) uncertainty arises when a machine learning model sees a test-time input that differs from its training data, and thus should not be predicted by the model. As ML is used for more safety-critical domains, the ability to handle out-of-distribution data are central in building open-world learning systems. In this talk, I will talk about challenges, methods, and opportunities on uncovering the unknowns of deep neural networks for reliable decision-making in an open world.
Speaker: Sharon Yixuan Li, University of Wisconsin-Madison
Bio: Sharon Yixuan Li is an Assistant Professor in the Department of Computer Sciences at the University of Wisconsin-Madison. Previously she worked as a postdoc research fellow in the Computer Science department at Stanford AI Lab (SAIL). She completed her Ph.D. from Cornell University in 2017, where she was advised by John E. Hopcroft. She has served as the Program Chair and founding organizer of the ICML Workshop on Uncertainty and Robustness in Deep Learning (UDL), Area Chair for NeurIPS, ICML, ICLR, and AAAI. Her broad research interests are in deep learning and machine learning. Her works explore, understand, and mitigate the many challenges where failure modes can naturally occur in deploying machine learning models in the open world. Research topics that she is currently focusing on include: (1) Out-of-distribution uncertainty estimation in deep learning; (2) Uncertainty-aware deep learning in healthcare and computer vision; and (3) Robustness to biases, data irregularity, and out-of-distribution generalization. She was named Forbes 30 Under 30 in Science, and 30 Under 30 Rising Stars in AI, and JP Morgan early-career outstanding faculty. Website: http://pages.cs.wisc.edu/~sharonli/
Invited Talk 1: Few-shot Network Anomaly Detection via Cross-network Meta-learning
Abstract: Network anomaly detection aims to find network elements (e.g., nodes, edges, subgraphs) with significantly different behaviors from the vast majority. It has a profound impact in a variety of applications ranging from finance, healthcare to social network analysis. Due to the unbearable labeling cost, existing methods are predominately developed in an unsupervised manner. Nonetheless, the anomalies they identify may turn out to be data noises or uninteresting data instances due to the lack of prior knowledge on the anomalies of interest. Hence, it is critical to investigate and develop few-shot learning for network anomaly detection. In real-world scenarios, few labeled anomalies are also easy to be accessed on similar networks from the same domain as of the target network, while most of the existing works omit to leverage them and merely focus on a single network. Taking advantage of this potential, in this work, we tackle the problem of few-shot network anomaly detection by (1) proposing a new family of graph neural networks -- Graph Deviation Networks (GDN) that can leverage a small number of labeled anomalies for enforcing statistically significant deviations between abnormal and normal nodes on a network; and (2) equipping the proposed GDN with a new cross-network meta-learning algorithm to realize few-shot network anomaly detection by transferring meta-knowledge from multiple auxiliary networks. Extensive evaluations demonstrate the efficacy of the proposed approach on few-shot or even one-shot network anomaly detection.
Speaker: Kaize Ding, Arizona State University
Invited Talk 2: Masked Contrastive Learning for Anomaly Detection
Abstract: Detecting anomalies is one fundamental aspect of a safety-critical software system, however, it remains a long-standing problem. Numerous branches of works have been proposed to alleviate the complication and have demonstrated their efficiencies. In particular, self-supervised learning based methods are spurring interest due to their capability of learning diverse representations without additional labels. Among self-supervised learning tactics, contrastive learning is one specific framework validating their superiority in various fields, including anomaly detection. However, the primary objective of contrastive learning is to learn task-agnostic features without any labels, which is not entirely suited to discern anomalies. In this paper, we propose a task-specific variant of contrastive learning named masked contrastive learning, which is more befitted for anomaly detection. Moreover, we propose a new inference method dubbed self-ensemble inference that further boosts performance by leveraging the ability learned through auxiliary self-supervision tasks. By combining our models, we can outperform previous state-of-the-art methods by a significant margin on various benchmark datasets.
Speaker: Hyunsoo Cho, Seoul National University
Invited Talk 3: Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning
Abstract: Anomaly detection with weakly supervised video-level labels is typically formulated as a multiple instance learning (MIL) problem, in which we aim to identify snippets containing abnormal events, with each video represented as a bag of video snippets. Although current methods show effective detection performance, their recognition of the positive instances, i.e., rare abnormal snippets in the abnormal videos, is largely biased by the dominant negative instances, especially when the abnormal events are subtle anomalies that exhibit only small differences compared with normal events. This issue is exacerbated in many methods that ignore important video temporal dependencies. To address this issue, we introduce a novel and theoretically sound method, named Robust Temporal Feature Magnitude learning (RTFM), which trains a feature magnitude learning function to effectively recognise the positive instances, substantially improving the robustness of the MIL approach to the negative instances from abnormal videos. RTFM also adapts dilated convolutions and self-attention mechanisms to capture long- and short-range temporal dependencies to learn the feature magnitude more faithfully. Extensive experiments show that the RTFM-enabled MIL model (i) outperforms several state-of-the-art methods by a large margin on four benchmark data sets (ShanghaiTech, UCF-Crime, XD-Violence and UCSD-Peds) and (ii) achieves significantly improved subtle anomaly discriminability and sample efficiency. Code is available at https://github.com/tianyu0207/RTFM
Speaker: Yu Tian, University of Adelaide