Many robots today are confined to operate in relatively simple, controlled environments. One reason for this is that current methods for processing visual data tend to break down when faced with occlusions, viewpoint changes, poor lighting, and other challenging but common situations that occur when robots are placed in the real world. I will show that we can train robots to handle these variations by modeling the causes behind visual appearance changes. If we model how the world changes over time, we can be robust to the types of changes that objects often undergo. I demonstrate this idea in the context of autonomous driving, and I show how we can use this idea to improve performance on three different tasks: velocity estimation, segmentation, and tracking with neural networks. By modeling the causes of appearance changes over time, we can make our methods more robust to a variety of challenging situations that commonly occur in the real-world, thus enabling robots to come out of the factory and into our lives.
David Held is a Post-doctoral Researcher at U.C. Berkeley working with Pieter Abbeel. He recently completed his Ph.D. in Computer Science at Stanford, doing research at the intersection of robotics, computer vision, and machine learning. His Ph.D. was co-advised by Sebastian Thrun and Silvio Savarese. David has also interned at Google, working on the self-driving car project. Before Stanford, he worked as a software developer for a startup company and was a researcher at the Weizmann Institute, working on building a robotic octopus. He received a B.S. in Mechanical Engineering at MIT in 2005, an M.S. in Mechanical Engineering at MIT 2007, and an M.S. in Computer Science at Stanford in 2012, for which he was awarded the Best Master’s Thesis Award from the Computer Science Department.
The big data revolution has profoundly changed, among many other things, how we perceive business, research, and application. However, in order to fully realize the potential of big data, certain computational and statistical challenges need to be addressed. In this talk, I will present my research in facilitating the deployment of machine learning methodologies and algorithms in big data applications. I will first present robust methods that are capable of accounting for uncertain or abnormal observations. Then I will present a generic regularization scheme that automatically extracts compact and informative representations from heterogeneous, multi-modal, multi-array, time-series, and structured data. Next, I will discuss two gradient algorithms that are computationally very efficient for our regularization scheme, and I will mention their theoretical convergence properties and computational requirements. Finally, I will present a distributed machine learning framework that allows us to process extremely large-scale datasets and models. I conclude my talk by sharing some future directions that I am and will be pursuing.
Yaoliang Yu is currently a research scientist affiliated with the center for machine learning and health, and the machine learning department of Carnegie Mellon University. He obtained his PhD (under Dale Schuurmans and Csaba Szepesvari) in computing science from University of Alberta (Canada, 2013), and he received the PhD Dissertation Award from the Canadian Artificial Intelligence Association in 2015.
Microsurgery ranks among the most challenging areas of surgical practice, requiring the manipulation of extremely delicate tissues by various micron scale maneuvers and the application of very small forces. Vitreoretinal surgery, as the most technically demanding field of ophthalmic surgery, treats disorders of the retina, vitreous body, and macula, such as retinal detachment, diabetic retinopathy, macular hole, and epiretinal membrane. Recent advancements in medical robotics have significant potential to address most of the challenges in vitreoretinal practice, and therefore to prevent traumas, lessen complications, minimize intra-operative surgeon effort, maximize surgeon comfort, and promote patient safety. In this talk, I will present the development of novel force-sensing tools and robot control methods to produce integrated assistive surgical systems that work in partnership with surgeons against the current limitations in microsurgery, specifically focusing on membrane peeling and vein cannulation tasks in retinal microsurgery. Integrating high sensitivity force sensing into the ophthalmic instruments enables precise quantitative monitoring of applied forces. Auditory feedback based upon the measured forces can inform (and warn) the surgeon quickly during the surgery and help prevent injury due to excessive forces. Using these tools on a robotic platform can attenuate hand tremor of the surgeon, which effectively promotes tool manipulation accuracy. In addition, based upon certain force signatures, the robotic system can actively guide the tool towards clinical targets, compensate any involuntary motion of the surgeon, or generate additional motion that will make the surgical task easier. I will present our latest experimental results for two distinct robotic platforms, the Steady Hand Robot and Micron, with the force-sensing ophthalmic instruments, which show significant performance improvement in artificial dry phantoms and ex-vivo biological tissues.
Berk Gonenc is a Ph.D. candidate in Mechanical Engineering at Johns Hopkins University. He received his M.S. degree in Mechanical Engineering from Washington State University Vancouver in 2011 and joined the Advanced Medical Instrumentation and Robotics Research Laboratory in Johns Hopkins University. He received his M.S.E. degree in Mechanical Engineering from Johns Hopkins University in 2014. His research is focused on developing smart instruments and robot systems for microsurgery.
In this talk, I will give a perspective on past and current trends in medical imaging particularly regarding the role of imaging in personalized medicine. I will then outline core technologies enabling the advancements, with specific focus on empirical and mechanistic modeling. In addition, I demonstrate some example clinical applications on how mechanistic and empirical models derived based on imaging are used for treatment planning and therapy outcome analysis. I conclude by providing a future outlook for the utilization of imaging in the healthcare continuum.
For over a decade surgical educators have called for objective, quantitative methods to measure surgical skill. To date, no satisfactory method exists that is simultaneously accurate, scalable, and generalizable. That is, a method whose scores correlate with patient outcomes, can scale to cope with 51 million annual surgeries in the United States, and generalize across the diversity surgical procedures or specialties. This talk will review the promising results of exploiting crowdsourcing techniques to meet this need. The talk will also survey the limitations of this approach, fundamental problems in establishing ground truth for surgical skill evaluation, and steps to exploit surgical robotics data. The talk will conclude by proposing some future robotic approaches that may obviate the need for surgeons to master complex technical skills in the first place.
Dr. Kowalewski completed his PhD in electrical engineering for “quantitative surgical skill evaluation” at the University of Washington’s Biorobotics lab. This work was recognized with a best doctoral candidate award at the American College of Surgeons AEI Consortium on Surgical Robotics and Simulation. He was also a research scientist at DARPA’s “Traumapod: Operating room of the future” project. He has helped commercialize his PhD work for quantitative skill evaluation hardware (Simulab Corp., Seattle, WA) and also pioneered the use of crowdsourcing for highvolume assessment of surgical skills and cofounded CSATS Inc, Seattle, WA to make these methods available to modern healthcare. This work has been published in JAMA Surgery and formally adopted by the American Urological Association for educational and certification needs. In 2012 he started the Medical Robotics and Devices Lab at the University of Minnesota, Mechanical Engineering department where he is currently an Assistant Professor.
This is the Fall 2016 Kick-Off Seminar, presenting an overview of LCSR, useful information, and an introduction to the faculty and labs.
While robots at the human size scale are generally composed of structures that are moved by a small set of actuators that shift materials or components with a well-defined shape, other principles for designing moving structures can control movement at the micron scale. For example, cells can move by disassembling parts of their rigid skeleton, or cytoskeleton, and reassembling new components in a different location. The structures that are disassembled and reassembled are often filaments that grow, shrink and form junctions between one another. Networks of rigid filaments serve as a cheap, reusable, movable scaffold that shapes and reshapes the cell.
Could we design synthetic materials to perform tasks of engineering interest at the micron scale? I’ll describe how we are using ideas from DNA nanotechnology to build synthetic filaments and how we can program where and when filaments assemble and disassemble and how they organize. We are able to use quantitative control over microscopic parameters, modeling and automated analysis to build increasingly sophisticated structures that can find, connect and move locations in the environment, form architectures and heal when damaged.
Bernhard Fuerst is a research engineer at the Engineering Research Center at Johns Hopkins University. He received his Bachelor’s degree in Biomedical Computer Science at the University for Medical Technology in Austria in 2009 and his Master’s degree in Biomedical Computing at the Technical University in Munich, Germany in 2011. During his studies he joined Siemens Corporate Research in Princeton to research biomechanical simulations for compensation of respiratory motion under Dr. Ali Kamen’s supervision, and Georgetown University to investigate techniques for meta-optimization using particle swarm optimizers under Dr. Kevin Cleary’s supervision. Since joining the Johns Hopkins University, he worked on establishing Dr. Nassir Navab’s research group to focus on robotic ultrasound, minimally invasive nuclear imaging, and bioelectric localization and navigation.