Architecture-Preserving Provable Repair of Deep Neural Networks
Explore a groundbreaking approach to repairing deep neural networks (DNNs) in this 21-minute conference talk from PLDI 2023. Delve into the concept of architecture-preserving V-polytope provable repair, which guarantees that repaired DNNs satisfy given specifications on an infinite set of points within a defined V-polytope. Learn how this method modifies DNN parameters without altering the architecture, supports various activation functions and layer types, and runs in polynomial time. Discover the implementation of this approach in the APRNN tool and compare its efficiency, scalability, and generalization to previous repair methods using MNIST, ImageNet, and ACAS Xu DNNs. Gain insights into this innovative technique that addresses the critical issue of incorrect behavior in DNNs, potentially mitigating disastrous real-world consequences.