2nd Workshop on Geometry and Machine Learning
Thursday, July 6, 2017 | Brisbane, Australia
Organized by Jinhui Xu and Jeff Phillips.
Machine learning concerns techniques that can learn from and make predictions on data. Algorithms for these tasks are built to explore the useful patterns in data, which usually can be stated in terms of geometry (e.g., problems in high dimensional feature space). Hence computational geometry can play a crucial and natural role in machine learning, espcecially since geometric algorithms often come with quality guaranteed solutions when dealing with the high-dimensional data. The Computational Geometry community has many researchers with the unique knowledge of relevant geometric tasks, which could be utilized to have a great impact on machine learning and other data related fields.

This workshop is intended to provide a forum for those working in the fields of computational geometry, machine learning and the various theoretical and algorithmic challenges to promote their interaction and blending. To this end, the workshop will consist of several invited talks that will serve as tutorials on a specific applications of geometric algorithms in machine learning. We hope such interaction will stimulate those working in both fields, and we can expect that a synergy can promote many new interesting geometric problems, concepts and ideas, which will contribute to open up new vistas in computational geometry communities.

This workshop is being held as part of CG Week 2017 (July 4-7, 2017 in Brisbane, Australia) which also includes the International Symposium on Computational Geometry (SoCG).

Invited Speakers
2:30-3:15 Yusu Wang (The Ohio State University) Metric denoising: A geometric perspective
3:15-4:00 Brian Lovell (The University of Queensland) Federated Model for Surveillance and Non-Cooperative CCTV Face Recognition
4:30-5:15 Anne Driemel (Eindhoven University of Technology) Algorithms and Data Structures in Spaces of Curves
5:15-6:00 Yufei Tao (Chinese University of Hong Kong) On the computational hardness of DBSCAN: From Static to Dynamic

Metric denoising: A geometric perspective
Yusu Wang
Associate Professor, The Ohio State University
Many geometric and topological methods rely on the metric information of input data. In this talk, I will explore three natural setups for modeling noise associated with input data/metric. Specifically, (i) the input data could already be embedded in a metric space, but the embedding may be noisy; or (ii) we are given a noisy discrete n-point metric as input which could include outliers; and we aim to remove these outliers and embed remaining points to a target metric space with low distortion; or (iii) the target metric is induced from an input graph which itself is noisy, leading to noisy graph metrics. I will describe the specific noise model in each setup and briefly discuss denoising approaches for each case to recover the metric information with theoretical guarantees. Our methods draw ideas from computational geometry, theoretical computer science and statistical modeling.

Bio: Yusu Wang is Professor of Computer Science and Engineering Department at the Ohio State University. She obtained her PhD degree from Duke University, where she received the Best PhD Dissertation Award at CS Dept., Duke U, in 2004. Before joining OSU in 2005, she was a post-doctoral fellow at Stanford University from 2004--2005. Yusu Wang works in the fields of Computational geometry and Computational topology. She is primarily interested in developing effective and theoretically justified algorithms for data / shape analysis using geometric and topological ideas and methods, and in applying them to practical domains, including computational biology, computer graphics and visualization. She received DOE Early Career Principal Investigator Award in 2006, NSF Career Award in 2008, and several best paper awards, including the Best Paper award in SIGSPATIAL 2015 and the Mark Fulk Best Student Paper award in COLT 2015.

Federated Model for Surveillance and Non-Cooperative CCTV Face Recognition
Brian Lovell
Professor, The University of Queensland
In this keynote I will describe our biometrics and surveillance research work on a huge surveillance project currently running with face recognition appliance nodes in Australia, UK, and Brazil. This fully operational system runs securely over the internet with edge processing to massively reduce bandwidth requirements and improve privacy. There is no need for a dedicated fibre network to connect all the high speed cameras - indeed wireless and mobile connectivity is a viable option. We have developed advanced low-resolution video face recognition technologies which work effectively and affordably in this unique architecture. The talk will describe this developing transcontinental surveillance system and how it is designed to be fully scalable to both national and international operation. A single cloud-based incident management backbone is accessible to all users from anywhere in the world. Along the way we will discuss the basics of robust CCTV -based video face recognition and the huge technical challenges of simultaneous pose, expression, illumination, obscuration, and motion blur compensation. We will also discuss recent work on robust face detection, land marking, and tracking to enable our systems to work on a crowd of people walking quickly past the cameras. The systems also perform cross-camera matching to measure queue lengths and estimates the gender and age of customers for retail applications. There will be live demonstrations of our systems in real-time operation including both mobile and wearable face recognition apps.

Bio: Brian C. Lovell was born in Brisbane, Australia in 1960. He received the BE in electrical engineering in 1982, the BSc in computer science in 1983, and the PhD in signal processing in 1991: all from the University of Queensland (UQ). Professor Lovell is Director of the Advanced Surveillance Group in the School of ITEE, UQ. He was President of the International Association for Pattern Recognition (IAPR) [2008-2010], and is Fellow of the IAPR, Senior Member of the IEEE, and voting member for Australia on the Governing Board of the IAPR. He is General Chair of the International Conference on Biometrics on the Gold Coast, Australia in 2018. He was Program Co-Chair of the International Conference of Pattern Recognition (ICPR2016) in CancĂșn Mexico, and was General Co-Chair of the IEEE International Conference on Image Processing in Melbourne, 2013 and Program Co-Chair of ICPR2008 in Tampa, Florida. His interests include non-cooperative Face Recognition, robust face detection, Biometrics, and Pattern Recognition. His work in biometrics and surveillance has won numerous international awards including the prestigious Best CCTV System at IFSEC2011, Birmingham, for Face in the Crowd recognition. He also won the government sponsored Asia Pacific ICT Trophy for Best R&D in the Asia Pacific region in Phuket, Thailand in 2011.

4:00-4:30 : Coffee Break

Algorithms and Data Structures in Spaces of Curves
Anne Driemel
Assistant Professor, Eindhoven University of Technology
Modern data analysis needs to be able to take sequences or time series as input. One can think of trajectories of moving objects, readings from sensors installed for monitoring, and access statistics of webpages or search trends. The Frechet distance measures the similarity of geometric curves, such as the ones mentioned above. There is a large body of research on the complexity of computing and approximating the Frechet distance, however it has a strong focus on computing single distances. This ignors the larger context in which the distance computation is usually invoked: clustering, nearest-neighbor searching and range searching. In my talk I will describe some recent work in this area.

Bio: Anne Driemel received her PhD from Utrecht University in the Netherlands. Her work focuses on the computational geometry of curve similarity computation. She is assistant professor in the group on data mining at the Eindhoven Technical University.

On the computational hardness of DBSCAN: From Static to Dynamic
Yufei Tao
Professor, Chinese University of Hong Kong
DBSCAN is a method proposed in 1996 for clustering multi-dimensional points, and has received extensive applications. In recent years, there has been major progress towards understanding its computational hardness. In this talk, we will discuss the current lower and upper bounds, covering (i) both the static and dynamic (including "insertion-only", and "insertions plus deletions") settings, and (ii) both the exact and approximate versions.

Bio: Yufei Tao is a Professor in the Department of Computer Science and Engineering, Chinese University of Hong Kong. He served as an associate editor of ACM Transactions on Database Systems (TODS) from 2008 to 2015, and of IEEE Transactions on Knowledge and Data Engineering (TKDE) from 2012 to 2014. He served as a PC co-chair of International Conference on Data Engineering (ICDE) 2014. He gave a keynote speech at International Conference on Database Theory (ICDT) 2016. He received two best-paper awards at SIGMOD (in 2013 and 2015, respectively), the title of ACM Distinguished Scientist in 2016, and a Google Faculty Research Award in 2017.

The previous version of this workshop was: Workshop on Geometry and Machine Learning.