Room 206 (2nd floor, badged access)
12 March 2020 - 10h30
Safety Certification for Deep Learning: Crazy Talk or Glimmer of Hope?
by Jyotirmoy Deshmukh from University of Southern California in Los Angeles,
Abstract: Deep Reinforcement Learning is a new and exciting technique most prominently used to showcase how DNN-based controllers can achieve tasks that were heretofore thought to be very difficult for computers; examples include, defeating top human players at computer games and strategy games, adding autonomous driving capabilities to cars, and making drones fly themselves. This has led to newfound optimism that DNNs may be eventually be used for safety-critical applications. Researchers with formal methods and computational logic sensibilities need to be prepared to answer the question: would you sit in a car driven by neural networks? How would you certify a safety-critical system using neural networks? This is an incredibly complex challenge problem. In this talk, we look at some baby steps on solving this grand challenge. We look at three different lines of investigation that the CPS-VIDA group at USC is pursuing: (1) training deep neural network controllers to do what they do, but while being safe, (2) designing local state-based rewards from high-level abstract logical specifications, and (3) sanity checking for convolutional neural networks for object detection that have aspirations of being used in a self-driving car.
Bio: Jyotirmoy V. Deshmukh (Jyo) is an assistant professor in the CS department at the University of Southern California in Los Angeles, USA. Before joining USC, Jyo worked as a Principal Research Engineer in Toyota Motors North America R&D. He got his Ph.D. degree from the University of Texas at Austin and was a post-doctoral fellow at the University of Pennsylvania. Jyo's research interest is the broad areas of analysis, design, verification and synthesis of cyber-physical systems (CPS), and his current research focuses on challenge problems at the intersection of CPS, AI and autonomy.
