Design of Robotics and Embedded systems, Analysis, and Modeling Seminar (DREAMS)
The Design of Robotics and Embedded systems, Analysis, and Modeling Seminar (DREAMS) occurs weekly on Mondays from 4.10-5.00 p.m. in 250 Sutardja Dai Hall.
Similar to last year, DREAMS has joined forces with the Control Theory Seminar and the CITRIS People and Robots Seminar CPAR.
The Design of Robotics and Embedded systems, Analysis, and Modeling Seminar topics are announced to the DREAMS list, which includes the chessworkshop workgroup, which includes the chesslocal workgroup.
Information on the seminar series might be useful for potential speakers. If you have any questions about DREAMS, please contact Markus N. Rabe. If you want to subscribe to our mailing list, please drop me a line.
Seminars from previous semesters can be found here.
| Ioannis Paschalidis || January 27, 2017
| Julie Shah || February 13, 2017
| Marjan Sirjani || February 24, 2017
| Rajeev Joshi || February 27, 2017
| Brian Lathrop || March 03, 2017
| Clark Barrett || March 06, 2017
| Sergey Levine || March 13, 2017
| Stanley Osher || March 20, 2017
| Tim Salcudean || April 03, 2017
| Parvez Ahammad || April 10, 2017
| Steven Shladover || April 17, 2017
| Necmiye Ozay || April 24, 2017
| Loris D'Antoni || May 12, 2017
| Mark Ho || May 12, 2017
Data-Driven Price-of-Anarchy Estimation in Transportation Networks
Jan 27, 2017,
10-11am, 380 Soda, Ioannis Paschalidis, Boston University.
Equilibrium modeling is common in a variety of fields such as game theory, transportation science, and systems biology. The inputs for these models, however, are often difficult to estimate, while their outputs, i.e., the equilibria they are meant to describe, are often directly observable. By combining ideas from inverse optimization with the theory of variational inequalities, we develop an efficient, data-driven technique for estimating the parameters of these models from observed equilibria. A distinguishing feature of our approach is that it supports both parametric and nonparametric estimation by leveraging ideas from statistical learning.
We apply this general framework to transportation networks. Using real traffic data from the Boston area, we estimate origin-destination flow demand matrices and the per-road cost (congestion) functions drivers implicitly use for route selection. Given this information, one can formulate and solve a system-optimum problem to identify socially optimal flows for the transportation network. The ratio of total latency under a user-optimal policy versus
a system-optimal policy is the so-called Price-of-Anarchy (POA), quantifying the efficiency loss of selfish actions compared to socially optimal ones. We find that POA can be quite substantial, sometimes exceeding 2, suggesting that there is scope for control actions to steer the equilibrium to a socially optimal one. We will discuss what some of these actions may be and how to prioritize interventions.
Yannis Paschalidis is a Professor of Electrical and Computer Engineering, Systems Engineering, and Biomedical Engineering at Boston University. He is the Director of the Center for Information and Systems Engineering (CISE). He obtained a Diploma (1991) from the National Technical University of Athens, Greece, and an M.S. (1993) and a Ph.D. (1996) from the Massachusetts Institute of Technology (MIT), all in Electrical Engineering and Computer Science. He has been at Boston University since 1996. His current research interests lie in the fields of systems and control, networks, applied probability, optimization, operations research, computational biology, medical informatics, and bioinformatics.
Prof. Paschalidis' work has been recognized with a CAREER award from the National Science Foundation, the second prize in the George E. Nicholson competition by INFORMS, and a finalist best paper award in the IEEE International Conference on Robotics and Automation (ICRA). His work on protein docking has been recognized for best performance in modeling selected protein-protein complexes against 64 other predictor groups. Work with students has won a best student paper award at the 9th Intl. Symposium of Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks, an IBM/IEEE Smarter Planet Challenge Award, and an IEEE Computer Society Crowd Sourcing Prize. He was an invited participant at the 2002 Frontiers of Engineering Symposium organized by the National Academy of Engineering, and at the 2014 National Academies Keck Futures Initiative (NAFKI) Conference. Prof. Paschalidis is a Fellow of the IEEE and the Editor-in-Chief of the IEEE Transactions on Control of Network Systems.
Enhancing Human Capability with Intelligent Machine Teammates
Feb 13, 2017,
4-5pm, 250 SDH, Julie Shah, MIT.
Every team has top performers -- people who excel at working in a team to find the right solutions in complex, difficult situations. These top performers include nurses who run hospital floors, emergency response teams, air traffic controllers, and factory line supervisors. While they may outperform the most sophisticated optimization and scheduling algorithms, they cannot often tell us how they do it. Similarly, even when a machine can do the job better than most of us, it can’t explain how. In this talk I share recent work investigating effective ways to blend the unique decision-making strengths of humans and machines. I discuss the development of computational models that enable machines to efficiently infer the mental state of human teammates and thereby collaborate with people in richer, more flexible ways. Our studies demonstrate statistically significant improvements in people’s performance on military, healthcare and manufacturing tasks, when aided by intelligent machine teammates.
Julie Shah is an Associate Professor in the Department of Aeronautics and Astronautics at MIT and leads the Interactive Robotics Group of the Computer Science and Artificial Intelligence Laboratory. In 2014, Shah was recognized by the National Science Foundation with a Faculty Early Career Development (CAREER) award and by MIT Technology Review on its 35 Innovators Under 35 list. Her work on industrial human-robot collaboration was also in Technology Review’s 2013 list of 10 Breakthrough Technologies. She has received international recognition in the form of best paper awards and nominations from the ACM/IEEE International Conference on Human-Robot Interaction, the American Institute of Aeronautics and Astronautics, the Human Factors and Ergonomics Society, the International Conference on Automated Planning and Scheduling, and the International Symposium on Robotics. Shah earned degrees in aeronautics and astronautics and in autonomous systems from MIT.
Can We Trust Self-Driving Cars? Adaptive Timed Actors for Building Dependable Cyberphysical Systems
Feb 24, 2017,
2-3pm, 540 Cory Hall, Marjan Sirjani, Mälardalen University.
In this presentation I will not talk about self-driving cars. I put the phrase in the title to catch your attention. I will talk about models, techniques and tools that can be used to build dependable cyberphysical systems (and hence be able to trust self-driving cars).
A family of actor-based languages are introduced to enable model driven development and provide a faithful and usable model for building distributed, asynchronous, and event-based systems with least effort. Network and computational delays, periodic events, and required deadlines can be expressed in the model. Model checking and simulation tools are built based on the formal semantics of the language. For deadlock-freedom and schedulability analysis special efficient techniques in state space exploration is proposed by exploiting the isolation of method execution in the model. I will show how these models can be used in safety assurance and performance evaluation of different systems, like Network on Chip architectures, sensor network applications, Traffic Control systems, and quadricopters. I show a general pattern in track-based traffic control systems, and a framework where self-adaptive actors are used to address self-adaptive traffic control systems.
Marjan Sirjani joined Malardalen University in June 2016 as a Professor and the Chair of the Software Engineering group. She is also a part-time Professor at School of Computer Science at Reykjavik University. Her main research interest is applying formal methods in Software Engineering. She works on modeling and verification of concurrent, distributed and cyberphysical systems.
Marjan and her research group are pioneers in building model checking tools, compositional verification theories, and state-space reduction techniques for actor-based models. Marjan has been the PC member and PC chair of several international conferences including SEFM, Coordination, FM, FMICS, ICFEM, MEMOCODE, and FSEN. Before joining academia as a full-time faculty she has been the managing director of Behin System Company for more than ten years, developing software and providing system services. Marjan served as the head of the Software Engineering Department of School of Electrical and Computer Engineering at the University of Tehran prior to joining the School of Computer Science at Reykjavik University in 2008.
She visited Ptolemy group at UC Berkeley as a Fulbright Scholar in 2015 and her research is currently focused on safety assurance and performance evaluation of autonomous cyberphysical systems, in which she is collaborating with Ptolemy group.
The Bugs that went to Mars and Terrorized Earth
Feb 27, 2017,
4-5pm, 250 SDH, Rajeev Joshi, NASA JPL.
Since its dramatic landing in Gale crater in August 2012, the Curiosity Rover has been busy exploring the surface of Mars, looking for evidence of past habitable environments. Having completed over 4 years on Mars, and with nearly 17 kms on its odometer, Curiosity has already made historic discoveries, finding evidence of an ancient freshwater streambed, organic molecules and other key ingredients necessary for life. Yet, in spite of its great successes, the mission has not been without a few hiccups. In this talk, we discuss the most significant of these: the Sol-200 anomaly, when the failure of a flash memory chip uncovered three latent software bugs that nearly killed the mission. We describe how the anomaly manifested itself, how recovery was achieved, and lessons learnt from the experience. The work described in this talk was carried out at Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
Rajeev Joshi is a Principal Engineer at the Lab for Reliable Software at NASA's Jet Propulsion Laboratory, where he works on building and applying tools based on formal methods to improve mission software reliability. He is also currently the Chief Engineer for Flight Software and Avionics Systems at JPL. He was a member of the Curiosity rover flight software development team, and, after landing, a member of the surface operations team, serving as data management chair and supporting anomaly investigations. For his work on Curiosity, he received two JPL Mariner Awards and the NASA Exceptional Achievement Medal. He holds a B.Tech in Computer Science from the Indian Institute of Technology, Bombay, and an MS/PhD (also in Computer Sciences) from the University of Texas at Austin. His previous employment includes 4 years at the DEC/Compaq/HP Systems Research Center (SRC) in Palo Alto, CA, and 2 years at AT&T Bell Labs in Murray Hill, NJ. He is an elected member (and current secretary) of IFIP Working Group 2.3 on Programming Methodology.
3+1 - An HMI Design Framework for Autonomous Vehicles
Mar 03, 2017,
2-3pm, 540 Cory, Brian Lathrop, Volkswagen of America, Electronics Research Lab.
While mode confusion and the ensuing human error that goes with it will likely have a significant impact on the safety of future AVs, the transitioning between modes and how those transitions are orchestrated via the vehicle’s HMI will be equally important. This is particularly relevant when the modes to which one is transitioning are not discrete states (e.g., on and off). That is, when transitions are put into the context of SAE Driving Automation Definitions it becomes clear that the human operator will transition into partial automation, conditional automation, and high automation. These variable states of the vehicle need to be communicated in a timely and clear manner, and the orchestration of the transitions between states needs to be effortless and exact.
Senior Principal Scientist Brian Lathrop is the Senior Principal Scientist for the Technology and Trend Scouting team at Volkswagen. In 2003 Brian received his Ph.D. in Cognitive Science from the University of California, Santa Cruz. In 2004 Brian joined VW and was responsible for human factors and usability testing activities for infotainment and driver assistance systems. In 2008 Brian became the senior manager of the HMI team at the ERL, responsible for defining the vision, roadmap, and overall strategy. He has led many projects focused on reinventing the vehicle cockpit of tomorrow, realizing advanced infotainment controls, futuristic displays, gaze and gesture-dependent interfaces, and HMI concepts for self-driving cars. In 2016 Brian joined the Technology and Trend Scouting, focused on transforming customer insights into user friendly products.
Automatic discovery and localization of tough bugs in large SoCs using formal-enhanced quick error detection
Mar 06, 2017,
4-5pm, 250 SDH, Clark Barrett, Stanford.
Quick error detection (QED) is an existing technique that transforms existing SoC test suites to improve coverage and reduce error detection latency. I will discuss two recent results which use formal methods to greatly enhance the power of QED. First, Symbolic QED is a method which uses bounded model checking to exhaustively search for short sequences of instructions which could cause QED checks to fail. Symbolic QED finds logic bugs and is applicable both to pre- and post-silicon designs. Second, Electrical QED is a technique that uses a small amount of additional hardware coupled with bounded model checking and QED to quickly localize post-silicon electrical bugs to specific design blocks and even to a handful of flip-flops within those design blocks. Both techniques were evaluated on the OpenSPARC T2 SoC with a wide variety of injected logic and electrical bugs. The new techniques automatically found and localized the injected bugs.
Clark Barrett is an associate professor (research) of computer science at Stanford University, with expertise in constraint solving and its applications to verification. His PhD dissertation introduced a novel approach to constraint solving now known as satisfiability modulo theories (SMT). His subsequent work on SMT has been recognized with a best paper award at DAC, an IBM Software Quality Innovation award, the Haifa Verification Conference award, and first-place honors at the SMT, CASC, and SyGuS competitions. He was also an early pioneer in the development of formal hardware verification: at Intel, he collaborated on a novel theorem prover used to verify key microprocessor properties; and at 0-in Design Automation (now part of Mentor Graphics), he helped build one of the first industrially successful assertion-based verification tool-sets for hardware.
Deep Robotic Learning
Mar 13, 2017,
4-5pm, 250 SDH, Sergey Levine, UC Berkeley.
Deep learning methods have provided us with remarkably powerful, flexible, and robust solutions in a wide range of passive perception areas: computer vision, speech recognition, and natural language processing. However, active decision making domains such as robotic control present a number of additional challenges, standard supervised learning methods do not extend readily to robotic decision making, where supervision is difficult to obtain. In this talk, I will discuss experimental results that hint at the potential of deep learning to transform robotic decision making and control, present a number of algorithms and models that can allow us to combine expressive, high-capacity deep models with reinforcement learning and optimal control, and describe some of our recent work on scaling up robotic learning through collective learning with multiple robots.
Sergey Levine received a BS and MS in Computer Science from Stanford University in 2009, and a Ph.D. in Computer Science from Stanford University in 2014. He joined the faculty of the Department of Electrical Engineering and Computer Sciences at UC Berkeley in fall 2016. His work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning algorithms. Applications of his work include autonomous robots and vehicles, as well as computer vision and graphics. His research includes developing algorithms for end-to-end training of deep neural network policies that combine perception and control, scalable algorithms for inverse reinforcement learning, deep reinforcement learning algorithms, and more.
Overcoming the Curse of Dimensionality for Hamilton-Jacobi equations with Applications to Control and Differential Games
Mar 20, 2017,
4-5pm, 250 SDH, Stanley Osher, UCLA.
It is well known that certain Hamilton-Jacobi partial differential equations (HJ PDE’s) play an important role in analyzing control theory and differential games. The cost of standard numerical algorithms for HJ PDE’s is exponential in the space dimension and time, with huge memory requirements. Here we propose and test methods for solving a large class of these problems without the use of grids or significant numerical approximation. We begin with the classical Hopf and Hopf-Lax formulas which enable us to solve state independent problems via variational methods originating in compressive sensing with remarkable results. We can evaluate the solution in 10^(-4) to 10^(-8) seconds per evaluation on a laptop. The method is embarrassingly parallel and has low memory requirements. Recently, with a slightly more complicated, but still embarrassingly parallel method, we have extended this in great generality to state dependent HJ equations, apparently, with the help of parallel computers, overcoming the curse of dimensionality for these problems. The term, “curse of dimensionality” was coined by Richard Bellman in 1957 when he did his classic work on dynamic optimization.
Stanley Osher is a Professor of Mathematics, Computer Science, Chemical Engineering and Electrical Engineering at UCLA. He is also an Associate Director of the NSF-funded Institute for Pure and Applied Mathematics at UCLA. He received his MS and PhD degrees in Mathematics from the Courant Institute of NYU. Before joining the faculty at UCLA in 1977, he taught at SUNY Stony Brook, becoming professor in 1975. He has received numerous academic honors and co-founded three successful companies, each based largely on his own (joint) research. Osher has been elected to the US National Academy of Science and the American Academy of Arts and Sciences. He was awarded the SIAM Pioneer Prize at the 2003 ICIAM conference and the Ralph E. Kleinman Prize in 2005. He was awarded honorary doctoral degrees by ENS Cachan, France, in 2006 and by Hong Kong Baptist University in 2009. He is a SIAM and AMS Fellow. He gave a one hour plenary address at the 2010 International Conference of Mathematicians. He also gave the John von Neumann Lecture at the SIAM 2013 annual meeting. He is a Thomson-Reuters highly cited researcher-among the top 1% from 2002-2016 in both Mathematics and Computer Science with an h index of 105. In recent years he gets cited approximately once per houron Google scholar. In 2014 he received the Carl Friedrich Gauss Prize from the International Mathematics Union-this is regarded as the highest prize in applied mathematics. In 2016 he received the William Benter prize from the City University of Hong Kong. His current interests involve information science, which includes optimization, image processing, compressed sensing and machine learning and applications of these techniques to the equations of physics, engineering and elsewhere. More recently he has been working on overcoming the "curse of dimensionality in control theory, differential games, and elsewhere.
Ultrasound and ultrasound-mediated image guidance for robot assisted surgery
Apr 03, 2017,
4-5pm, 250 SDH, Tim Salcudean, University of British Columbia.
Medical robotic systems present a great opportunity for integrating imaging with surgical navigation. Indeed, the instruments are localized and tracked in real time with respect to the camera view, so once registered to the patient, imaging can be used to display anatomy and pathology with respect to the robot camera and the instruments.
We present our approaches to integrating ultrasound and magnetic resonance imaging with the da Vinci medical robotic system for prostate surgery. We will summarize our calibration and registration techniques and our experience from a first patient study (N=27) in which this system was used.
Integration of ultrasound with medical robots enables new intra-operative ultrasound-based tissue characterization. We present our work in two areas: quantitative elastography to measure tissue shear storage and loss moduli (“objective palpation”), and photoacoustic imaging to measure blood oxygenation.
Tim Salcudean received his bachelor’s and master’s from McGill University and the doctorate from U.C. Berkeley, all in Electrical Engineering. From 1986 to 1989, he was a Research Staff Member in the robotics group at the IBM T.J. Watson Research Center. He then joined the Department of Electrical and Computer Engineering at the University of British Columbia, Vancouver, Canada, where he holds a Canada Research Chair and the Laszlo Chair in Biomedical Engineering.
Professor Salcudean’s research contributions have been in the areas of medical imaging, medical robotics, simulation and virtual environments, haptics, teleoperation and optimization-based design. Several companies have licensed his technology and his gland-contouring software for prostate cancer radiotherapy has become the standard of care in British Columbia, and has been used in well over 2000 patients. Prof. Salcudean has been a co-organizer of several research symposia and has served as a Technical and Senior Editor of the IEEE Transactions on Robotics and Automation. He is a Fellow of MICCAI, the IEEE, and the Canadian Academy of Engineering.
ML for web application security: Reconciling two competing philosophies
Apr 10, 2017,
4-5pm, 250 SDH, Parvez Ahammad, Instart Logic Inc..
The idea of using machine learning (ML) to solve problems in security
domains is almost three decades old. Security attacks broadly fall
into two classes - those which occur as a result of known
vulnerabilities; and those that arise from previously unknown classes
of bugs. The risk and prevalence of both these categories of attacks
are increasing and this in turn has motivated researchers to turn to
more methodical approaches to deal with the amount of data necessary
to defend against these attacks. The application of machine learning
to security has been growing in prominence as a seemingly natural fit
to automatically prioritize and classify events. However, not only is
the jury is still out on how well ML solves security problems, there
is a real tension between largely probabilistic machine learning
approaches to security and the security mindset of "correct by
In this talk, I will provide a historical overview of academic
publications from the last decade that applied ML in security domains.
I will explore the generalized system designs, underlying assumptions,
measurements, and use cases in active research. I will also share a
taxonomy on ML paradigms and security domains for future exploration
along with an agenda detailing open and upcoming challenges for
applying ML in security. Along the way, I will present some recent
results from our work on enabling enterprise web application security
at scale by combining the "correct by construction" approaches with
ML-driven approaches for better coverage against attacks.
Parvez Ahammad leads the data science and machine learning efforts at
Instart Logic. His group is focused on creating data-driven
algorithms, and innovative product features that optimize and secure
web application delivery at scale. He has applied machine learning in
a variety of domains, most recently to computational neuroscience (at
HHMI-Janelia), web application delivery, and web application security.
Along the way, he has mentored data scientists, built teams, and has
had to grapple with issues like explainability and interpretability of
ML systems, insufficient amount of labeled data, scalability, ethics,
and adversaries who target ML models. Parvez holds a PhD in electrical
engineering and computer sciences from UC Berkeley, with an emphasis
in computer vision and machine learning.
Road Vehicle Automation: History, Opportunities and Challenges
Apr 17, 2017,
4-5pm, 250 SDH, Steven Shladover, .
The concept of road vehicle automation has captured the imagination of the general public and many in the technical community recently. Most do not recognize how long a history this concept has had, so this talk goes back to 1939 and works forward from then to give a historical context for prior developments in this field. The opportunities for improving road transportation with automation are explained, along with the severe scientific and technological challenges that remain to be addressed. This talk explains why it is going to be a lot harder and take a lot longer to replace human drivers for the majority of their driving functions than most people currently believe. However, much progress can be achieved in the nearer term by focusing on more limited applications of automated driving, within carefully constrained operational design domains.
Dr. Steven Shladover has been researching road vehicle automation systems for more than forty years, beginning with his masters and doctoral theses at M.I.T. He is the Program Manager, Mobility at the California PATH Program of the Institute of Transportation Studies of the University of California at Berkeley. He led PATH’s pioneering research on automated highway systems, including its participation in the National Automated Highway Systems Consortium from 1994-98, and has continued research on fully and partially automated vehicle systems since then. This work has included definition of operating concepts, modeling of automated system operations and benefits, and design, development and testing of full-scale prototype vehicle systems. His target applications have included cooperative adaptive cruise control, automated truck platoons, automated buses and fully-automated vehicles in an automated highway system.
Dr. Shladover joined the PATH Program in 1989, after eleven years at Systems Control, Inc. and Systems Control Technology, Inc., where he led the company’s efforts in transportation systems engineering and computer-aided control engineering software products. He chaired the Transportation Research Board Committee on Intelligent Transportation Systems from 2004-2010, and currently chairs the TRB Committee on Vehicle-Highway Automation. He was the chairman of the Advanced Vehicle Control and Safety Systems Committee of the Intelligent Transportation Society of America from its founding in 1991 until 1997. Dr. Shladover leads the U.S. delegation to ISO/TC204/WG14, which is developing international standards for “vehicle-roadway warning and control systems”.
Control Synthesis for Large Collections of Dynamical Systems with Counting Constraints
Apr 24, 2017,
4-5pm, 250 SDH, Necmiye Ozay, University of Michigan.
ABSTRACT: Can we control a swarm of systems and give guarantees on their collective behavior? In this talk I will discuss an instance of this problem: given a large nearly homogeneous collection of dynamical systems and a novel class of safety constraints, called counting constraints, how to synthesize a controller that guarantees the satisfaction of these constraints. Counting constraints impose restrictions on the number of systems that are in a particular mode or in a given region of the state-space over time. I will present an approach for synthesizing correct-by-construction controllers to enforce such constraints. Our approach exploits the structure of the problem, the permutation invariance of dynamics due to homogeneity and the permutation invariance of counting constraints, to achieve massive scalability. I will discuss several extensions and potential applications of this approach. Finally, I will illustrate the approach on the problem of coordinating a large collection of thermostatically controlled loads while ensuring a bound on the number of loads that are extracting power from the electricity grid at any given time.
BIOGRAPHY: Necmiye Ozay received the B.S. degree from Bogazici University, Istanbul in 2004, the M.S. degree from the Pennsylvania State University, University Park in 2006 and the Ph.D. degree from Northeastern University, Boston in 2010, all in electrical engineering. She was a postdoctoral scholar at California Institute of Technology, Pasadena between 2010 and 2013. She is currently an assistant professor of Electrical Engineering and Computer Science, at the University of Michigan, Ann Arbor. Her research interests include dynamical systems, control, optimization, formal methods with applications in cyber-physical systems, system identification, verification and validation, and autonomy. Dr. Ozay is the recipient of a DARPA Young Faculty Award in 2014 and an NSF CAREER Award, a NASA Early Career Faculty Award and a DARPA Director’s Fellowship in 2016.
Adventures in program repair
May 12, 2017,
10-11am, 540 Cory, Loris D'Antoni, University of Wisconsin-Madison.
Programmers constantly face errors that are confusing and hard to fix. In particular, inexperience programmers, who cannot address commonly occurring errors, have to resort to online help-forums for finding corrections to their buggy programs.
I will present three ideas that leverage advances in program synthesis to automatically assist unskilled programmers that face commonly occurring errors.
First, I will talk about "emulative program repair", a technique for automatically learning how to fix errors directly from programmers. Given examples of bug fixes from real programmers, we synthesize "rules" that generalize such bug fixes and use them to automatically repair programs that contains similar bugs.
Second, I will talk about "repair via direct manipulation", a technique that allows the programmer to express what a repaired program should do by directly manipulating intermediate runtime values of a program. Using the manipulated values, we automatically synthesize programs compliant with the user intent.
Third, I will talk about "program repair under uncertainty", where program inputs are drawn from a probability distribution. I will show how program that do not satisfy a given probabilistic postcondition can be efficiently repaired by combining techniques from program synthesis and computational learning theory.
Loris D'Antoni is an Assistant Professor in the MadPL (Madison Programming Languages and Software Engineering) Group at the University of Wisconsin-Madison. He received his PhD from the University of Pennsylvania in 2015, where he worked under the supervision of Rajeev Alur. His dissertation "Programming using Automata and Transducers" won the The Morris and Dorothy Rubinoff Award. Loris's research interests lie in formal methods and program synthesis, with applications to networking, personalized education, and data science. Loris is currently investigating how formal methods can be used to provide personalized feedback to computer science students, automatically program networks, and remove bias from machine learning classifiers.
How People Intentionally Teach Agents in Interactive Settings
May 12, 2017,
1-2pm, 540 Cory, Mark Ho, Brown University.
People intuitively teach other people, animals, and even machines. Psychologically, teaching involves multiple interacting processes including modeling the learning agent, predicting its future behavior, and deciding on how teaching actions can facilitate learning. This raises questions about social cognition and interaction, as well as questions of how to design systems that can best leverage people’s capacity for pedagogy.
In this talk, I will discuss two lines of research that seek to answer how people teach in interactive settings. The first is on teaching by evaluative feedback, in which a person provides rewards and punishments to teach a learner a behavior. Teaching rewards are often conceived as incentives that shape the behavior of a learner. However, I will present results showing that this does not come naturally to people. Instead, we find that people use rewards/punishments to communicate whether a course of action is correct, even in online, interactive settings where this strategy is visibly counterproductive. The second line of work is on teaching by demonstration, in which a person takes actions to intentionally show how to perform a task. When intentionally teaching by demonstration, we find that people modify their behavior in ways that are suboptimal for doing the activity, but optimal for communicating their intent based on context.
Together, these findings indicate that teaching rewards and demonstrations differ from those expressed non-communicatively. Moreover, this work suggests that in contexts where autonomous agents must collaborate with and adapt to humans, it will be especially important to design systems that can distinguish and benefit from intentional teaching.
Mark Ho is a computational cognitive scientist and Ph.D. student at Brown University. His work examines social cognition from a reinforcement learning perspective, focusing on how people communicate and teach through non-verbal interaction. In addition to better understanding human cognition, his work formulating human communication as computational theories provides a foundation for developing machines that can more naturally interact with and learn from people.