NCCR Symposium Day 1 - The interplay of dynamical systems, neural networks and control

Event Type
Conference
Date
-
Location
EPFL
Room
EPFL BM 5202
Zoom: https://ethz.zoom.us/j/65954817501

 

Flyer for NCCR Symposium

What is this symposium about? 

This symposium will feature an outstanding line-up of world-wide experts in the field who will present their results and answer questions in a panel discussion. Each part of the workshop will be followed by a reception concluding the day, and individual meetings between the speakers and researchers will be arrange during the speakers' stay.

You can rewatch the talks and panel discussion from the day on our YouTube channel:

Symposium program

14:00 - 14:20: Welcome and introduction by Giancarlo Ferrari Trecate

14:20 - 15:00: "Stable adaptation and learning in large dynamical networks" by Jean-Jacques Slotine

15:00 - 15:40: "Incorporating dynamical system and control structure into neural networks " by Zico Kolter

15:40 - 16:10: Coffee break

16:10 - 16:50: "Learning to Fly" by Davide Scaramuzza

16:50 - 17:30: "Building Certifiably Safe and Correct Large-scale Autonomous Systems " by Chuchu Fan

17:30 - 18:30: Panel discussion

18:30 - 20:00: Apéro

Invited experts

Talk Abstract: The human brain still largely outperforms robotic algorithms in most tasks, using computational elements 7 orders of magnitude slower than their artificial counterparts. Similarly, current large scale machine learning algorithms require millions of examples and close proximity to power plants, compared to the brain's few examples and 20W consumption. We study how modern nonlinear systems tools, such as contraction analysis, virtual dynamical systems, and adaptive nonlinear control can yield quantifiable insights about collective computation, adaptation, and learning in large dynamical networks. 

Stable concurrent learning and control of dynamical systems is the subject of adaptive nonlinear control. When multiple parameter choices are consistent with the data (be it for insufficient richness of the task or aggressive overparametrization), stable Riemannian adaptation laws can be designed to implicitly regularize the learned model. Thus, local geometry imposed during learning may be used to select parameter vectors for desired properties such as sparsity. The results can also be systematically applied to predictors for dynamical systems. Stable implicit sparse regularization can be exploited as well to select relevant dynamic models out of plausible physically-based candidates. 

In optimization, most elementary results on gradient descent based on convexity of a time-invariant cost can be replaced by much more general results based on contraction. Semi-contraction of a natural gradient in some metric implies convergence to a global minimum, and furthermore that all global minima are path-connected. Adaptive controllers or predictors can also be used in transfer learning or sim2real contexts, where an optimizer has been carefully learned for a nominal system, but needs to remain effective in real-time in the presence of significant but structured variations in parameters. 

Finally, a key aspect of contraction tools is that they also suggest systematic mechanisms to build progressively more refined networks and novel algorithms through stable accumulation of functional building blocks and motifs. 

Biography: Jean-Jacques Slotine is Professor of Mechanical Engineering and Information Sciences, Professor of Brain and Cognitive Sciences, and Director of the Nonlinear Systems Laboratory. He received his Ph.D. from the Massachusetts Institute of Technology in 1983, at age 23. After working at Bell Labs in the computer research department, he joined the faculty at MIT in 1984. Professor Slotine teaches and conducts research in the areas of dynamical systems, robotics, control theory, computational neuroscience, and systems biology. He has been a Distinguished Faculty at Google AI since 2019. One of the most cited researchers in systems science, he was a member of the French National Science Council from 1997 to 2002,  a member of Singapore’s A*STAR SigN Advisory Board from 2007 to 2010, and has been a member of the Scientific Advisory Board of the Italian Institute of Technology since 2010. 

  • Prof. Zico Kolter (Carnegie Mellon) - Incorporating dynamical system and control structure into neural networks 

Talk Abstract: Neural networks have become a key tool for the modeling and control of dynamical systems.  However, typically these networks serve as black boxes, without guarantees of stability, controllability, or other factors that play a foundational role in the analysis of dynamical systems.  In this talk, I will discuss several approaches to integrating structure into neural networks to make them better suited to modeling and controlling dynamical systems.  This kind of structure includes integrating optimization problems, explicit model predictive control laws, or stability guarantees through Lyapunov projections.  Finally, I will also discuss recent work on leveraging advances in neural network verification to formally verify the stability of dynamical systems, offering an insight into how neural networks may also aid in more traditional analysis of such systems. 

Biography: Zico Kolter is an Associate Professor in the Computer Science Department at Carnegie Mellon University, and also serves as chief scientist of AI research for the Bosch Center for Artificial Intelligence. His work spans the intersection of machine learning and optimization, with a large focus on developing more robust and rigorous methods in deep learning. In addition, he has worked in a number of application areas, highlighted by work on sustainability and smart energy systems. He is a recipient of the DARPA Young Faculty Award, a Sloan Fellowship, and best paper awards at NeurIPS, ICML (honorable mention), AISTATS (test of time), IJCAI, KDD, and PESGM. 

Talk Abstract: I will summarize our latest research in learning deep sensorimotor policies for agile vision-based quadrotor flight. Learning sensorimotor policies represents a holistic approach that is more resilient to noisy sensory observations and imperfect world models. However, training robust policies requires a large amount of data. I will show that simulation data is enough to train policies that transfer to the real world without fine-tuning. We achieve one-shot sim-to-real transfer through the appropriate abstraction of sensory observations and control commands. I will show that these learned policies enable autonomous quadrotors to fly faster and more robustly than before, using only onboard cameras and computation. Applications include acrobatics, high-speed navigation in the wild, and autonomous drone racing. 

Biography: Davide Scaramuzza is a Professor of Robotics and Perception at the University of Zurich, where he does research at the intersection of robotics, computer vision, and machine learning. His goal is to enable autonomous, agile navigation of micro drones using both standard and neuromorphic event-based cameras. He pioneered autonomous, vision-based navigation of drones, which inspired the navigation algorithm of the NASA Mars helicopter. He has served as a consultant for the United Nations on topics such as disaster response and disarmament, as well as the Fukushima Action Plan on Nuclear Safety. He won many prestigious awards for his research contributions, such as a European-Research-Council Consolidator grant, the IEEE Robotics and Automation Society Early Career Award, an SNF-ERC Starting Grant, a Google Research Award, a Facebook Distinguished Faculty Research Award, two NASA TechBrief Awards, and several paper awards. In 2015, he co-founded Zurich-Eye, today Facebook Zurich, which developed the world-leading virtual-reality headset, Oculus Quest, which sold over 10 million units. Many aspects of his research have been prominently featured in broader media, such as The New York Times, The Economist, Forbes, BBC News, and Discovery Channel. 

Prof. Chuchu Fan (MIT) - Building Certifiably Safe and Correct Large-scale Autonomous Systems 

Talk Abstract: The introduction of machine learning (ML) and artificial intelligence (AI) creates unprecedented opportunities for achieving full autonomy. However, learning-based methods in building autonomous systems can be extremely brittle in practice and are not designed to be verifiable. In this talk, I will present several of our recent efforts that combine ML with formal methods and control theory to enable the design of provably dependable and safe autonomous systems. I will introduce our techniques to generate safety certificates and certified decision and control for complex autonomous systems, even when the systems have a large number of agents, follow nonlinear and nonholonomic dynamics, and need to satisfy high-level specifications. 

Biography: Chuchu Fan an Assistant Professor in the Department of Aeronautics and Astronautics at MIT. Before that, she was a postdoc researcher at Caltech and got her Ph.D. from the Electrical and Computer Engineering Department at the University of Illinois at Urbana-Champaign in 2019. She earned her bachelor’s degree from Tsinghua University, Department of Automation. Her group at MIT works on using rigorous mathematics including formal methods, machine learning, and control theory for the design, analysis, and verification of safe autonomous systems. Chuchu’s dissertation work “Formal methods for safe autonomy” won the ACM Doctoral Dissertation Award in 2020.