CS&E Colloquium: Testing Techniques for Learning Enabled Systems

The computer science colloquium takes place on Mondays from 11:15 a.m. - 12:15 p.m. 

This week's speaker, Sanjai Rayadurgam (CS&E), will be giving a talk titled "Testing Techniques for Learning Enabled Systems".

Abstract

Machine Learning enabled systems are being increasingly adopted in diverse domains where safety, security and reliability are critical concerns.  The ML techniques enabling these novel applications are typically data-driven and come with inherent risks that make system behavior difficult to predict and reason about. Thus, verification and validation of these systems pose challenges that are not easily addressed by established methods for "traditional" software driven systems. Specifically in the context of systems that require a high degree of assurance, such as autonomous flight, self-driving vehicles and medical devices, new ML specific assurance techniques grounded in well-established V&V principles are needed. In this talk, we will cover some recent and ongoing work on testing techniques for learning enabled systems by our research group.  We will also briefly review an overall assurance framework for such systems and see where the proposed techniques fit. 

Bio

Sanjai Rayadurgam is a Research Project Specialist in the Department of Computer Science and Engineering, and is the Director of the University of Minnesota Software Engineering Center. His research interests are in software testing, formal analysis and requirements engineering, with particular focus on application to high-assurance systems. Prior to his academic career he worked in the medical devices industry, developing systems engineering tools, and performing verification and validation of implantable cardiac devices. As part of his doctoral research at the University of Minnesota he developed techniques to create test cases meeting stringent test coverage criteria automatically from software behavioral models. He has co-authored several research papers and articles in software engineering.