Thursday, September 10, 2026
11:00 AM – 12:00 PM CDT
Abstractions for Scalable Verification of AI-enabled Cyber-Physical Systems
About this event
A Center for Autonomy seminar at UT's Oden Institute on formally verifying safety in systems that embed neural networks for perception and control — introducing interval abstractions that make verification of large networks and camera pipelines tractable. Free and open.
Proving safety for systems that learned their behaviour
Machine learning components now sit inside safety-critical cyber-physical systems across transportation, energy and medicine — handling perception, control and decision-making. That creates a hard problem: conventional formal verification does not scale to networks of realistic size, and it has no natural way to reason about a camera feeding a perception pipeline.
What the talk covers
Dr. Pavithra Prabhakar presents a formal approach to verifying closed-loop systems that combine dynamical models of physical processes with neural-network perception and control. She treats two scenarios: controllers implemented directly as neural networks, and perception pipelines that pair camera models with networks.
The core contribution is abstraction. Two data structures are introduced — Interval Neural Networks, which give abstract representations of network behaviour, and Interval Images, symbolic abstractions over whole sets of images — together with abstraction-refinement algorithms that search efficiently for abstractions small enough to prove safety. Experimental results target scalability on large systems. The talk also covers verification of evolving networks and ongoing work on stability analysis, refinement checking and compositional analysis.
The speaker
Prabhakar is Professor of Computer Science and the Cleve Moler and MathWorks Endowed Chair in Mathematical and Engineering Software at the University of New Mexico, previously at Kansas State. Her research is formal methods for AI-enabled autonomous, cyber-physical and robotic systems, with aerospace, automotive and agricultural applications. She recently served as a Program Director at the NSF's CISE Directorate.
Practical notes
Free and open, held in POB 6.304 in the Peter O'Donnell Jr. Building on the UT campus. No registration is listed. Best suited to people with a background in formal methods, control, or ML systems engineering.
Organized by
A UT Austin research institute working across computational science, mathematics, and engineering. Its public seminar series brings visiting researchers to campus for technical talks, frequently covering machine learning, foundation models, and scientific computing.
All events by Oden Institute for Computational Engineering and Sciences →