Navigation auf uzh.ch
Navigation auf uzh.ch
08.10.2020 - Open-set Classification: Know When You Don't Know
Speaker: Prof. Dr. Manuel Günther
Host: Prof. Dr. Manuel Günther
The classification of object categories from images has a long history in research and there exist many different approaches to achieve such a classification. But only with the deep learning boom, automatic classification of objects from images has matured into an age where image samples from thousands of different classes can be classified correctly with high probability and confidence. However, classifiers still have difficulties when a sample from an unknown class is presented. Imagine a classifier that has learned to distinguish pictures of apples, pears and bananas. When you show a picture of an apple to the classifier, it will likely classify it as an apple even if had never seen this particular breed of apple. But when you show an image of an orange to the classifier, it has no other chance than choosing one of the three classes it knows: apple, pear or banana, and it will usually do this with high confidence, though the correct answer would be: I don't know. In this talk I will present the concepts of open-set recognition and some ideas on how to teach classifiers to say: I don't know. I will include some extensions of well-known machine learning methods, and introduce our novel and exciting technique for incorporating this ability into deep learning models.
As an application, I will show how this technique can be used for recognizing faces on watchlists from surveillance cameras while ignoring innocent passersby and other artifacts that arise in real surveillance applications.
Already in younger ages, Manuel Günther attended a high school specialized on math and science in Erfurt, Germany, one of the few schools in Germany that offered Computer Science as a major subject. Consequently, he obtained a Master's degree in Computer Science (Dipl.-Inf.) from the Technical University of Ilmenau, Germany, in 2006. Between 2006 and 2011, he worked as a PhD student at the Ruhr-University Bochum, Germany, under the supervision of late PD. Dr. Rolf Würtz. His thesis was about statistical extensions of Gabor wavelet-based facial image processing and he obtained his PhD degree (Dr.-Ing.) from the University of Ilmenau in 2012. From 2012 to 2015, he was working as a postdoctoral researcher in the Biometrics group at the Idiap Research Institute in Martigny, Switzerland, under the supervision of Prof. Dr. Sébastien Marcel.
Since then he is actively contributing to the open-source signal processing and machine learning library Bob (http://www.idiap.ch/software/bob), we he particularly implemented the Biometric Recognition packages, which he presented as a hands-on tutorial in the Biometrics Summer School in Kuala Lumpur, Malaysia, in 2016 and at the International Joint Conference on Biometrics in 2017. Between 2015 and 2018, he worked as a Research Associate in Prof. Dr. Terrance E. Boult's Vision and Security Technology lab at the University of Colorado Colorado Springs, USA, where he was researching deep learning techniques for the classification of facial attributes and open-set classification, particularly for open-set face recognition. He was leading the Face Recognition Evaluation in Mobile Environment held in collaboration with the International Conference on Biometrics in 2013 and the Unconstrained Face Detection and Open Set Recognition Challenges held in collaboration with the International Joint Conference on Biometrics in 2017. At the end of 2018, he moved back to Germany where he undertook an industry excursion as a research engineer in the trinamiX GmbH in Ludwigshafen.
Since July 2020, Manuel Günther is employed as an Assistant Professor in the Department of Informatics at the University of Zurich, Switzerland, where he leads the Artificial Intelligence and Machine Learning group. There, he is continuing his research on open-set classification and face recognition. Also, he is member of the Digital Society Initiative (DSI) where he is collaborating with other disciplines.
22.10.2020 - Global ERP and Local Workarounds: The Value of Non-Compliance
Speaker: Prof. Dr. Robert M. Davison
Host: Prof. Dr. Gerhard Schwabe
This presentation reports on an exploratory case study to investigate how warehouse employees work around an ERP software that cannot be used as designed due to work practices required by local conditions. The context involves the local Hong Kong operations of a global retailer of home textiles. Our 29 interviews at the site provide many perspectives about how an inadequate information system failed to support necessary work practices and how employees at the site responded by creating a feral IT system that helped them pursue their business responsibilities and objectives. We draw on a compliance view of technology use to suggest that unreflective compliance can be counterproductive; paradoxically, reflective non-compliance may bring greater benefit to both the organization and its customers.
Prof. Dr. Robert Davison is a Professor of Information Systems at the City University of Hong Kong. His research focuses on the use and misuse of information systems, especially with respect to problem solving, guanxi formation and knowledge management, in Chinese organisations. He has published over 200 articles in a wide variety of our premier journals and conferences. He is particularly known for his scholarship in the domain of action research. He primarily teaches MSc and MBA students in the areas of IT consulting, Knowledge Management and Global Information Systems. Within the AIS, Robert chaired the research ethics committee for many years. He currently chairs the IFIP’s WG 9.4 (Social Implications of Computing in Developing Countries) and is the Editor-in-Chief of both the Information Systems Journal and the Electronic Journal of Information Systems in Developing Countries. Robert travels extensively, seeking to understand how people in different contexts and cultures make sense of their lives with IS. Professionally, he seeks to enhance the inclusion of scholars from the global south within our community. To this end, he frequently travels in developing countries where he offers research seminars and workshops, engaging with local PhD students and scholars. As a researcher and as an editor, he champions local and indigenous perspectives.
05.11.2020 - Distributed Denial-of-Service in 2020 - Why we Still Care
Speaker: Prof. Sven Dietrich, Ph.D.
Host: Prof. Dr. Burkhard Stiller
In this 20-year retrospective, we discuss some of the challenges of dealing with Distributed Denial-of-Service (DDoS) attacks from its origins in 1999 to recent attacks. We describe first architectures of DDoS agents, the challenges of DDoS agent/bot forensics, the variety of topologies and command-and-control mechanisms for botnets over the years, the different victim populations from scientific workstations to IoT devices, as well as future (i.e. next-generation) and current Internet design considerations. We also illustrate how these attacks can still impact our infrastructure today and what we should consider defense mechanisms at the host and network levels.
Prof. Sven Dietrich, PhD is Professor in the Computer Science Department at Hunter College, City University of New York (CUNY). U.S.A., where he started in August 2020, and has also been affiliated with the PhD program in Computer Science at the CUNY Graduate Center since 2015. Prior joining CUNY Hunter, Prof. Dietrich was Associate Professor in the Mathematics and Computer Science Department at CUNY John Jay College of Criminal Justice, Assistant Professor in Computer Science at Stevens Institute of Technology, a Senior Member of the Technical Staff at Carnegie Mellon University Software Engineering Institute and CERT Research, and a Senior Security Architect at the NASA Goddard Space Flight Center.
Prof. Dietrich’s research has focused on network security, especially on the analysis of Distributed Denial-of-Dervice attacks, botnets, and the mitigation of such attacks, formal verification of security protocols, applied cryptography, software security, malware, and the ethics of computer security research. He has organized various security conferences as Program Chair (most recently IEEE CS SADFE 2020), Steering Committee member (DIMVA, SADFE), and Program Committee member. He has served on the IEEE Computer Society Board of Governors, and also as the Technical Activities Chair there.
Prof. Dietrich holds a Doctor of Arts in Mathematics, a MS in Mathematics, and a BS in Computer Science and Mathematics from Adelphi University, Long Island, U.S.A.
19.11.2020 - Eye Movement-based Biometric Identification
Speaker: Prof. Dr. Lena Jäger
Host: Prof. Dr. Lena Jäger
Human eye movements have been shown to exhibit individual characteristics that are relatively stable over time. Hence, it has been proposed to use eye movements for biometric identification. Although subsequent work has demonstrated the promise of eye movements as behavioral biometric characteristic, the main drawback of the proposed methods is that the time-to-identification is several orders of magnitude above the time needed by competing biometric technologies such as face recognition, fingerprints, or iris recognition.
I will present the DeepEyedentification model which in contrast to all previous methods, does not operate on engineered features but rather computes a latent representation of an eye gaze sequence by training a deep convolutional network on the raw eye tracking signal. I will show that this approach outperforms the best existing approaches by one order of magnitude in terms of identification accuracy and by two orders of magnitude in terms of the duration of the eye tracking recording needed to identify a user. Furthermore, I will present an extension of the DeepEyedentification network that allows the detection of presentation attacks: The model detects replay attacks by processing both a controlled but randomized stimulus and the ocular response to this stimulus. Finally, I will evaluate the model's performance on eye tracking data with varying spatial and temporal resolution.
After having completed an MA in Chinese Studies at the University of Freiburg, Germany, the Tongji-University Shanghai, China, and the University Paris 7 Denis-Diderot, France, Lena Jäger studied Clinical and Experimental Linguistics at the University of Potsdam, Germany. In her PhD work, she investigated the cognitive mechanisms that underlie human language processing. During her PhD in Cognitive Science in (University of Potsdam, Germany, 2015), she studied Computer Science and later joined the Machine Learning Research group at the University of Potsdam as a postdoctoral researcher. Prof. Dr. Lena Jäger joined the Department of Computational Linguistics of the University of Zurich in July 2020.
26.11.2020 - Interactive Visual Data Analysis
Speaker: Prof. Dr. Jürgen Bernard
Host: Prof. Dr. Jürgen Bernard
With interactive visual data analysis (IVDA), I refer to the combination of Information Visualization, Visual Analytics, Human-Computer Interaction, and Machine Learning. IVDA is a user-centered approach to data science, following the principle to foster the involvement of humans in the data analysis process by providing interactive visual interfaces. IVDA combines the strength of humans and machines in an iterative and incremental data analysis process and is useful to tackle a series of data science challenges. Examples for data-oriented challenges are heterogeneous data, dirty data, uncertain data, or unlabeled data. Important model-oriented challenges include data preprocessing, model building, model quality assessment, or model explanation. IVDA enables personalized data science solutions tailored to the information need of individual users. Along these lines, IVDA also makes machine learning applicable for larger user groups beyond experts in data science.
After an introduction to IVDA, this talk further motivates the human-centered and interactive perspective of data science using two of my research branches as examples. In the first example, I will introduce Visual-Interactive Labeling, a user-centered paradigm to solve data labeling problems - without labeled training data, no supervised machine learning (e.g., classification) can be applied. With the second example, I will demonstrate how visual analytics solutions can help medical experts to explore patient histories in large and complex electronic health records.
Prof. Dr. Jürgen Bernard is Assistant Professor at UZH and head of the Interactive Visual Data Analysis research group. Jürgen was postdoctoral research fellow at the University of British Columbia (UBC), Vancouver. He studied Computer Sciences with focus on Applied Visual Computing at the TU Darmstadt and received his PhD Degree in 2015, when he was with the Fraunhofer Institute for Computer Graphics Research IGD. Jürgen started his first Post-doc phase at TU Darmstadt at Interactive Graphics Systems Group (Computer Science), leading his own research group (Visual-Interactive Machine Learning). During this time Jürgen was awarded with the Dirk Bartz Price (Visual Computing in Medicine) and the Hugo-Geiger Price (Excellent Dissertations in the Fraunhofer Society).
Jürgen's research interest is on Visual Analytics and Information Visualization and includes the characterization, design, and evaluation of visual-interactive interfaces to combine the strengths of both humans and algorithms in interactive machine learning and interactive data science applications. Important application domains so far revolved around climate and Earth observation, digital libraries, human motion analysis, music classification, sports data analysis, stock chart analysis, as well as medical and patient-related research in particular.
03.12.2020 - The Relational Data Borg is Learning
Speaker: Prof. Dr. Dan Olteanu
Host: Prof. Dr. Dan Olteanu
As we witness the data science revolution, each research community legitimately reflects on its relevance and place in this new landscape. The database research community has at least three reasons to feel empowered by this revolution. This has to do with the pervasiveness of relational data in data science, the widespread need for efficient data processing, and the new processing challenges posed by data science workloads beyond the classical database workloads. The first two aforementioned reasons are widely acknowledged as core to the community’s raison d’être. The third reason explains the longevity of relational database management systems success: Whenever a new promising data-centric technology surfaces, research is under way to show that it can be captured naturally by variations or extensions of the existing relational techniques. Like the Star Trek’s Borg Collective co-opting technology and knowledge of alien species, the Relational Data Borg assimilates ideas and applications from connex fields to adapt to new requirements and become ever more powerful and versatile. Unlike the former, the latter moves fast, has great skin complexion, and is reasonably happy. Resistance is futile in either case.
In this talk, I will make the case for a first-principles approach to machine learning over relational databases that guided recent development in database systems and theory. This includes theoretical development on the algebraic and combinatorial structure of relational data processing. It also includes systems development on compilation for hybrid database and learning workloads and on computation sharing across aggregates in learning-specific batches. Such development can dramatically boost the performance of machine learning.
Dan Olteanu has recently become Professor for Big Data Science at the University of Zurich after spending over 12 years at the University of Oxford. Over the last two decades, he has published in the areas of database systems, database theory, and AI, contributing to XML query processing, incomplete information and probabilistic databases, factorised databases, in-database machine learning, incremental maintenance for analytics, and the commercial systems LogicBlox and relationalAI. He co-authored the book « Probabilistic Databases » (2011). He served or is serving as associate editor for PVLDB (2012, 2020), IEEE TKDE (2013-2015), ACM TODS (2018-), and the SIGMOD Record database principles column (2019-). He also served among others as PC vice chair for SIGMOD 2017 and will serve as PC chair for ICDT 2022. He is the recipient of the ICDT 2019 best paper award, SIGMOD 2018 Distinguished PC member award, an ERC Consolidator grant (2016), and an Oxford Outstanding Teaching award (2009).
