iLunch Talk (Nov 6) Deep Learning with LIDAR and its Vulnerabilities

What: Deep Learning with LIDAR and its Vulnerabilities 

When: Friday November 6, 12 noon – 1:00 pm

Where: Zoom Meeting ID: 899 1138 2631  Password: 206904

Who: Theodore Nowak
Affiliation: Data Scientist within the Data Science and Analytics group at Pacific Northwest

National Laboratory (PNNL).

Abstract: As Autonomous Vehicles and other Robotic systems become increasingly ubiquitous, so too do the Light Intensity Detection And Ranging (LIDAR) sensors that form the backbone of their perceptual systems. While historically these sensors have been primarily used to perform

Simultaneous Localization And Mapping (SLAM), increasingly many works have sought to
transfer advances in Deep Learning (DL) to this domain. As these new DL models are increasingly applied in real-world, safety-critical applications, questions about their safety and robustness arise. In this talk I will explore the growing body of research studying Machine Learning (ML) on point sets and LIDAR data, and the vulnerabilities therein. Concretely, I will discuss these models, Adversarial Machine Learning (AML), the difficulties confronted when generalizing AML attacks to physical, real-world scenarios, and our recent work identifying a novel vulnerability in this space.

Bio: Ted Nowak is a Data Scientist within the Data Science and Analytics group at Pacific Northwest National Laboratory (PNNL). Prior to that, he received a M.Sc. in Robotics at the University of Michigan (2018), worked as an Engineer in Research at the University of Michigan (2015-2016), and received a B.Sc.E. in Electrical Engineering from Case Western Reserve University (2015). His research interests lie in extending and verifying ML on non-traditional data types to support scientific understanding.

Host: School of Computing and Information Science, University of Maine

For questions: salimeh.yasaei@maine.edu