The First Workshop on Machine Learning, Optimization and Control
MLOC is a forum for leading researchers and industry experts on all aspects of machine intelligence and optimization, including machine learning techniques, pattern recognition, data mining, optimization theory and control techniques. In the context of this symposium, machine learning encompasses works on concurrent deep learning techniques, their associated optimization and control techniques. Given the rise of deep learning and big data analysis, the first MLOC is particularly interested in work that addresses new optimization techniques, and control tools that attempt to improve machine learning performance in big data analytics, and work towards improved industry intelligent systems.
Specific topics of interest include (but are not limited to):
- Deep learning techniques and applications
- Big data analysis
- Robust face recognition
- Data driven control
- Cyber Security
- Clustering and classification techniques
- Randomised algorithm
- Data mining and knowledge discovery
- Computer vision and image understanding
- Prof. Lei Zhang (张磊), The Hong Kong Polytechnic University
- Prof. Wanquan Liu (刘万泉), Curtin University
- Prof. Yongsheng Ou (欧勇盛), Shenzhen Institutes of Advanced Technology, CAS
- Prof. Aiguo Wu (吴爱国), Harbin Institute of Technology, Shenzhen
- Prof. Chao Xu (许超), Zhejiang University
- Zhiyang Wang
- Shichao Zhou (周世超), email@example.com
9:30 First Keynote
- Title: TBA (on deep learning, optimization and control in general review)
- Spekaer: Prof. Xiangchu Feng, Xidian University
10:15 Questions and Discussions
10:45 Second Keynote
- Title: Implementing the ADMM to big datasets: A case study of LASSO
- Speaker: Prof. Xiaoming Yuan, Hongkong University
11:30 Questions and Discussions
14:00 Third Keynote
- Title: Machine learning via Wasserstein statistical manifold
- Speaker: Prof. Wuchen Li, UCLA
14:45 Questions and Discussions
15:15 Fourth Keynote
- Title: TBA
- Prof. Lei Zhang, The Hong Kong Polytechnic University
16:00 Questions and Discussions
- 9:00 Title: TBA (on dictionary learning) | Speaker: Yong Xu, Harbin Institute of Technology (Shenzhen)
- 9:30 人工智能中的优化问题研究 | Lingchen Kong, Beijing Jiaotong University
- 10:00 Title: Cardiac image quantification and motion analysis based on deep neural network | Wufeng Xue, Shenzhen University
- 10:30 Title: TBA | Daoqiang Zhang (Tony), NUAA
- 11:00 Title: Computer image and image understanding | Wei Xie, South China University of Technology
- 11:30 会议结束
- 12:00 午餐
Title: Machine learning via Wasserstein statistical manifold
Abstract: In this talk, I start with reviewing several primal-dual structures in optimal transport (Wasserstein metric). Based on it, I will introduce the Wasserstein natural gradient in parametric statistical models. We pull back the L2-Wasserstein metric tensor in probability density space to parameter space, under which the parameter space become a Riemannian manifold. The gradient and Hamiltonian flows in parameter space are derived. When parameterized densities lie in 1D, we show that the induced metric tensor and gradient flow establish explicit formulas. Examples are presented to demonstrate its effectiveness in several machine learning problems.
浙江大学玉泉校区智能系统与控制研究所二楼资料室, 9 a.m., July 2, 2018
Prof. Ryan Loxton, full professor in the School of Electrical Engineering, Computing, and Mathematics at Curtin University, Australia.
Abstract: Switched systems operate by switching among various different modes. Determining the optimal times at which the mode switches should occur is a fundamental problem in systems and control, with particular importance to the numerical solution of optimal control problems. This talk will discuss the switching time optimization problem for two classes of switched systems: those with time-dependent switching conditions (where the switches are directly controllable), and those with state-dependent switching conditions (where the switches occur when the system hits certain switching surfaces in the state space). It is widely believed that standard numerical optimization techniques struggle when applied to switching time optimization problems. In this talk we present new results showing that this challenge is over-stated; contrary to popular belief, switching times can in fact be optimized effectively using standard optimization methods. We verify this with a numerical example involving a switched system model for the production of 1,3-propanediol, an industrial polymer used in paints, adhesives, and lubricants.
Bio of the speaker: Ryan Loxton is a full professor in the School of Electrical Engineering, Computing, and Mathematics at Curtin University, Australia. His research focuses on developing new mathematical techniques to optimize complex processes in a wide range of applications such as mining, oil and gas, agriculture, and industrial process control. Ryan’s work has been recognized with several high-profile awards, including two prestigious, highly competitive fellowships from the Australian Research Council and the 2014 West Australian Young Scientist of the Year Award. A passionate advocate for industry engagement, Ryan has led many industry-funded research projects with companies such as Woodside Energy, Linkforce, Roy Hill Iron Ore, Vekta Automation, and Global Grain Handling Solutions. His mathematical algorithms underpin the Quantum software system (developed by Onesun Pty Ltd) for tracking, executing, and optimizing maintenance shutdowns in the resources sector. This technology was the winner of the 2017 South32 Designing for Excellence Innovation Award. Ryan is an Associate Editor for the Journal of Industrial and Management Optimization and has published over 70 papers in international journals and conference proceedings.
Prof. Chao Xu is invited to deliver an aera talk (SS04: New Development of Smart Devices for Structural Control | 结构控制的智能装置新进展) at the 7th WCSCM, the World Conference on Structural Control and Monitoring (WCSCM). The title of Prof. Xu's talk is Infrastructure Cyber-Care: Challenges to Cyber-Systems, Robotics and Dada Analytics.
WCSCM is a premier leading conference, under the auspices of the International Association for Structural Control and Monitoring (IACSM). The WCSCM, held every four years, is aiming at promoting advanced structural control and monitoring technology for a variety of civil, mechanical, aerospace and energy systems. The precedent conferences have been held in Pasadena - USA (1994), Kyoto - Japan (1998), Como - Italy (2002), La Jolla - USA (2006), Tokyo - Japan (2010) and Barcelona - Spain (2014).The new edition of the WCSCM, 7WCSCM, will be hosted by Harbin Institute of Technology in July 2018. The conference will provide international research community a platform to contribute to the state of the art in such multidisciplinary scientific and engineering environment with new results, fresh ideas and future perspectives.
An Optimal Control Approach to Deep Learning
浙江大学玉泉校区教九101演讲厅, 9:30 a.m., June 25, 2018
Dr. Qianxiao Li, Research Scientist at the Institute of High Performance Computing, A*STAR, Singapore and an adjunct assistant professor in the department of mathematics, National University of Singapore
Abstract: In this talk, we discuss a new approach to study the algorithmic and theoretical aspects of deep learning. In particular, the optimization of deep neural networks is recast as an optimal control problem, which is a classical problem that originates from the calculus of variations. Based on this viewpoint, we investigate the development of novel algorithms, as well as theoretical insights on generalization bounds of neural networks.
Bio of the speaker: Qianxiao Li is a research Scientist at the Institute of High Performance Computing, A*STAR, Singapore and an adjunct assistant professor in the department of mathematics, National University of Singapore. He graduated with a BA in Mathematics from University of Cambridge in 2010, and a PhD in applied mathematics from Princeton University in 2016. His main research interests include the theory and algorithms for deep learning, stochastic optimization and applications ofmachine learning to the physical sciences.
- 报告人：祁振强 研究员，中国运载火箭技术研究院北京航天自动控制研究所
个人简历：祁振强，工学博士，研究员，现任中国航天科技集团有限公司中国运载火箭技术研究院北京航天自动控制研究所技术发展处处长、研究院核高基重大科技专项办公室副主任，国防973项目总体组组长，国家国防科技工业局国防基础科研计划信息与控制专家组专家。担任中英先进控制系统技术联合实验室中方主任、学术委员会委员，曾任IEEE ICMC 2014程序委员会主席，IEEE CGNCC 2016程序委员会委员；是《宇航学报》、《控制与信息技术》编委和多个学术期刊审稿人。从事航天运载器导航制导控制、智能自主控制等技术研究。完成3个国家航天装备型号任务和国防973等多项科研项目研究，获得国家技术发明一等奖、国防技术发明二等奖、军队科技进步二等奖等科技奖项7项，获得授权专利33项，获颁首次探月工程有功人员奖章。主持推进国际合作，建成中英先进控制系统技术联合实验室，获批航天先进控制技术国际联合研究中心国家国际科技合作基地。
Visualizing Real World Flows
Professor Steve Wereley, Purdue University, Mechanical Engineering
- 18:30, May 13, 2018
Particle Image Velocimetry (PIV) is a laboratory tool that many fluid mechanics professionals use to measure flows in their labs where experimental conditions can be perfectly controlled. However, quite often scientists need to know about flows in the real world. One example where fluid mechanics professionals must work in the field under non-ideal conditions is that of oil spilling into the environment. While the size of many oil spills is known because a fixed volume of oil has spilled out of a damage oil tanker, for many others, the size of the spill is not. For example, the Deepwater Horizon oil spill in the USA in 2010 occurred when the oil company lost control of an oil well and oil poured unrestrainedly into the Gulf of Mexico for more than 3 months (see picture below). In order to assess the amount of damage caused by the oil spill, Professor Wereley used quantitative flow visualization to estimate the size of the oil spill. This approach was shown to be much more accurate than other methods that were used to measure the size of the oil spill.
Abstract: Over last decade, small-size unmanned aerial vehicles (UAVs) have received unprecedented research interests and created extensive applications as well as market opportunities. Currently dominant UAV platforms, such as quadrotors and hexarotors, although exhibits great maneuverability (i.e. vertical takeoff and landing, hovering capability), is inherently energy inefficient by its flight mechanics. This drawback has increasingly limited their application in range- and/or endurance-demanding tasks such as surveying and mapping. Motivated by this, hybrid aerial vehicles such as tail-sitters, tilt-rotors, tilt-props, or dual-propulsion systems can transform between multirotor mode and fixed-wing mode, thus inheriting both benefits. Though having been actively explored in the aviation history, low-cost, small-size hybrid UAVs with increased intelligence and autonomy still poses a grand challenge. In this talk, I will present the development of a portable hybrid vertical takeoff and landing (VTOL) UAV. From a system point of view, three topics will be covered: design, modeling and control. Through the design, implementation and intensive tests, I will show how recent advances in low cost actuator (e.g. motor, propeller, ESCs), computation units and sensors have enabled the development of such small-size hybrid VTOL UAVs, and the arisen opportunities and challenges.
Bio: Dr. Fu Zhang received the B.E. (with honor) from the Department of Automation, University of Science and Technology of China, Hefei, China, in 2011. Then he studied at Department of Mechanical Engineering, the University of California, with full scholarship. He received the Ph.D. degree there in 2015, focusing on the dynamics modeling and control of ultra-high precision machining systems and high performance micro-scale rate integrating gyros funded by DARPA. Dr. Fu Zhang is now a research assistant professor with the Department of Electronic and Computer Engineering at the Hong Kong University of Science and Technology (HKUST). His current research interests include the dynamics, control and navigation of unmanned aerial vehicles (UAVs), deep reinforcement learning and optimization.
- 浙江大学工程力学系 学士
- 清华大学工程力学系 硕士
- 明尼苏达大学航空工程与力学系 博士
Nonlinear Control of Transport PDE-ODE Interconnections
Marie Sklodowska-Curie Fellow
Dep. of Production Eng. & Managem.
Technical University of Crete
Chania, Greece 73100
Abstract: Numerous physical processes are described by transport PDE-ODE interconnections. LTI systems with constant input delays is perhaps the most elementary class in this category, where the transport speed is constant and the boundary of the spatial domain is fixed, besides the plant being linear. For this class of systems, predictor feedback is now a well-known delay-compensating control design tool. The situation becomes dramatically more complex when, in addition to the ODE being nonlinear, the speed of propagation or the boundary of the domain is a nonlinear function of the overall infinite-dimensional state (i.e., of the PDE or the ODE state) of the system. For such interconnections, I will present predictor-feedback design ideas, which I will then illustrate with several application examples, including, traffic systems (where the transport speed is a nonlinear function of the PDE state), extruders for 3D printing (giving rise to a system with ODE state-dependent moving boundary), and metal rolling (where the transport speed is a nonlinear function of the ODE state).
Nikolaos Bekiaris-Liberis received the Ph.D. degree in Aerospace Engineering from the University of California, San Diego, in 2013. From 2013 to 2014 he was a postdoctoral researcher at the University of California, Berkeley and from 2014 to 2017 he was a research associate and adjunct professor at Technical University of Crete, Greece. Dr. Bekiaris-Liberis is currently a Marie Sklodowska-Curie Fellow at the Dynamic Systems & Simulation Laboratory, Technical University of Crete. He has coauthored the SIAM book Nonlinear Control under Nonconstant Delays. His interests are in delay systems, distributed parameter systems, nonlinear control, and their applications.
Dr. Bekiaris-Liberis was a finalist for the student best paper award at the 2010 ASME Dynamic Systems and Control Conference and at the 2013 IEEE Conference on Decision and Control. He received the Chancellor’s Dissertation Medal in Engineering from the University of California, San Diego, in 2014. Dr. Bekiaris-Liberis received the best paper award in the 2015 International Conference on Mobile Ubiquitous Computing, Systems, Services and Technologies. He is the recipient of a 2017 Marie Sklodowska-Curie Individual Fellowship Grant.