PostDoc Candidate Talk: Sherlock - A Tool For Verification Of Neural Network Feedback Systems

PostDoc Candidate Talk: Sherlock - A Tool For Verification Of Neural Network Feedback Systems
Thu, February 27, 2020 @ 10:30am EST
University of Pennsylvania
220 S 33rd Street
Towne Building Room 315
Philadelphia, Pennsylvania 19104
Speaker
Souradeep Dutta
University of British Columbia
Abstract

Neural networks have become ubiquitous when it comes to learning enabled cyber-physical systems, like autonomous cars and closed loop medical devices.  Such safety critical applications, lead to a compelling use case for formal verification approaches.  In this talk I will be presenting  an approach for verification of neural network controllers for closed loop dynamical systems.  Given a neural network and a set of possible inputs to the network described by polyhedral constraints, the aim would be to compute a safe over-approximation of the set of possible output values.  I would present an efficient range estimation algorithm that iterates between an expensive global combinatorial search using mixed-integer linear programming problems, and a relatively inexpensive local optimization that repeatedly seeks a local optimum of the function represented by the network.  Next, we show how this can be combined with tools like Flow*, to compute reach sets for systems where the dynamics are modeled as an ODE.

Speaker Bio

I am Souradeep Dutta, a 4th year PhD candidate at the Dept of Electrical Computer and Energy Engineering, University of Colorado at Boulder. I am a part of the CUPLV research group, where my adviser is Sriram Sankaranarayanan.