Stanford Seminar - Recent Progress in Verifying Neural Networks, Zico Kolter

Explore recent advancements in verifying neural networks in this Stanford seminar featuring Zico Kolter, Associate Professor at Carnegie Mellon University. Delve into the challenges of guaranteeing network output properties for specific input classes, a crucial aspect of validating robustness and safety in neural networks. Learn about the significant progress made in this complex field, with recent methods achieving verification speeds thousands of times faster than generic solvers. Discover the team's award-winning approach at the Verification of Neural Networks Competition (VNNCOMP) 2021. Gain insights into the verification problem, linear networks, off-the-shelf solvers, and practical applications. Examine the importance of robustness, potential security flaws, and the broader implications for deep learning validation.

Introduction
Overview
Recent examples
The bottom line
Verifying deep learning
The problem of deep networks
Linear networks
Offtheshelf solvers
Questions
Validation
Linear Programming
Branch Inbound
In practice
Does robustness matter
Security flaws
Why doesnt anyone care
Conclusion