COMPUTER SCIENCE RESEARCH WEEK JANUARY 2020
Professor Lothar Thiele, ETH Zurich
Assistant Professor Pan Jia, University of Hong Kong
COM1 Level 2
SR1, COM1-02-06
This is a distinguished talk as part of the NUS Computer Science Research Week 2020 https://researchweek.comp.nus.edu.sg/
10:00 - 11:30 Compressed Sensing and Generative Models - Eric Price
Abstract:
The goal of compressed sensing is make use of image structure to estimate an image from a small number of linear measurements. The structure is typically represented by sparsity in a well-chosen basis. We show how to achieve guarantees similar to standard compressed sensing but without employing sparsity at all -- instead, we suppose that vectors lie near the range of a generative model G: R^k -> R^n. Our main theorem here is that, if G is L-Lipschitz, then roughly O(k log L) random Gaussian measurements suffice; this is O(k d log n) for typical d-layer neural networks.
The above result describes how to use a model to recover a signal from noisy data. But if the data is noisy, how can we learn the generative model in the first place? The second part of my talk will describe how to incorporate the measurement process in generative adversarial network (GAN) training. Even if the noisy data does not uniquely identify the non-noisy signal, the _distribution_ of noisy data may still uniquely identify the distribution of non-noisy signals.
This talk is based on joint works with Ashish Bora, Ajil Jalal, and Alex Dimakis.
Biodata:
Eric Price is an assistant professor in the Department of Computer Science at UT Austin, where he studies how algorithms can produce more accurate results with less data. He received a Ph.D. in computer science from MIT in 2013. Eric's research was featured in Technology Review's TR10 list of 10 breakthrough technologies of 2012, his thesis received a George M. Sprowls award for best doctoral thesis in computer science at MIT, and he has received an NSF CAREER award. Two themes of his research are adaptivity, where initial data can guide future data collection, and signal structure, where a structural assumption can yield provable improvements in space or sample complexity.
13:00 - 14:30 Internet of Things - The Quest for Dependability - Lothar Thiele
Abstract:
If visions and forecasts of industry come true then we will be soon surrounded by billions of interconnected embedded devices. We will interact with them in a cyber-human symbiosis, they will not only observe us but also our environment, and they will be part of many visible and ubiquitous objects around us. We have the legitimate expectation that the individual devices as well as the overall system behaves in a reliable, predictable and trustworthy manner.
Besides, there are many application domains where we rely on a correct and fault-free system behavior. We expect trustworthy results from sensing, computation, communication and actuation due to economic importance or even catastrophic consequences if the overall system is not working correctly, e.g., in industrial automation, distributed control of energy systems, surveillance, medical applications, or early warning scenarios in the context of building safety or environmental catastrophes. Finally, trustworthiness and reliability are mandatory for the societal acceptance of human-cyber interaction and cooperation.
It will be argued that we need novel architectural concepts, an associated design process and validations strategies to satisfy the strongly conflicting requirements and associated design challenges of platforms for the Internet of Things: Handle at the same time limited available resources, adaptive run-time behavior, and predictability. These challenges concern all components and functions of an IoT system, e.g., information extraction from global data, local decision making, computation, storage, wireless communication, energy management, energy harvesting, sensors, sensor interfaces, and actuation. The focus of the presentation is on new models and methods as well as examples from various fields in environmental monitoring.
Biodata:
Lothar Thiele joined ETH Zurich, Switzerland, as a full Professor of Computer Engineering, in 1994. His research interests include models, methods and software tools for the design of embedded systems, internet of things, cyberphysical systems, sensor networks, embedded software and bioinspired optimization techniques.
Lothar Thiele is associate editor of INTEGRATION - the VLSI Journal, Journal of Signal Processing Systems, IEEE Transaction on Industrial Informatics, Journal of Systems Architecture, IEEE Transactions on Evolutionary Computation, Journal of Real-Time Systems, ACM Transactions on Sensor Networks, ACM Transactions on Cyberphysical Systems, and ACM Transaction on Internet of Things.
In 1986 he received the "Dissertation Award" of the Technical University of Munich, in 1987, the "Outstanding Young Author Award" of the IEEE Circuits and Systems Society, in 1988, the Browder J. Thompson Memorial Award of the IEEE, and in 2000-2001, the "IBM Faculty Partnership Award". In 2004, he joined the German Academy of Sciences Leopoldina. In 2005, he was the recipient of the Honorary Blaise Pascal Chair of University Leiden, The Netherlands. Since 2009 he is a member of the Foundation Board of Hasler Foundation, Switzerland. Since 2010, he is a member of the Academia Europaea. In 2013, he joined the National Research Council of the Swiss National Science Foundation SNF. Lothar Thiele received the "EDAA Lifetime Achievement Award" in 2015. Since 2017, Lothar Thiele is Associate Vice President of ETH Zurich for Digital Transformation.
15:30 - 17:00 Autonomous robotic systems by combining control and learning - Pan Jia
Abstract:
Thanks to the progress of control and machine learning techniques, robots nowadays are more intelligent than their predecessors 30 years ago. Advanced control techniques have enabled industrials robots to be faster, stronger, and more accurately than human workers in repetitive, structured tasks, though they are still difficult to solve everyday tasks in unstructured scenarios. Machine learning techniques have brought a revolution in perception and decision making, which allows a robot to explore new situations smartly according to its experience. However, its performance in many robotics tasks in terms of accuracy and robustness is far from being satisfactory. Control or machine learning alone is not enough for building an intelligent robot. An open problem in robotics is thus about how to combine control and machine learning in the most appropriate way in order to solve the bottleneck of applying robotics techniques in everyday life. In this talk, we will introduce our solutions to this non-trivial challenge using two challenging robotic tasks as examples: one is the autonomous robotic navigation in the dense pedestrian crowd, and the other is the deformable object manipulation for cloth assembly and suturing.
Biodata:
Pan Jia is currently an assistant professor in the Department of Computer Science at the University of Hong Kong. Prior to joining HKU, he was an assistant professor at the Mechanical Department at the City University of Hong Kong and a postdoc in the EECS department at the University of California at Berkeley. He received his BEng in Control Engineering from Tsinghua University in 2008, MEng in Pattern Recognition from the Chinese Academy of Sciences, and Ph.D. in Computer Science from the University of North Carolina at Chapel Hill. His research interests are robotic control and learning with a special focus on the development of autonomous robotic systems for navigation and manipulation.