Architecture-Preserving Provable Repair of Deep Neural Networks
summary
Presents an algorithm that modifies only network parameters to yield architecture-preserving fixes that are provably correct on entire V-polytopes, encoding convex sets by vertices and guaranteeing the specification over the infinite interior. Repair restricts symbolic variables to one layer and concretizes others to produce a quadratic-free linear formula solved by an LP, running in polynomial time and supporting ReLU-style piecewise-linear activations, fully-connected, convolutional, pooling and residual layers. Tool APRNN scales to ImageNet and MNIST, yields 23% and 61% generalization gains in reported cases, and repairs ACAS Xu properties in ~100s, outperforming prior non-architecture-preserving methods.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:19DNN repair challenges
- 3:28Limitations of repair methods
- 4:35APRON overview
- 5:17Semantics and repair approach
- 9:19Pointwise repair results
- 12:18Evaluating large networks
- 13:22Polytop high-dimensional results
- 14:53Safety properties in aircraft control
- 15:13Conclusion and future work
- 15:40Q&A
related talks
42mTowards Verification of General Neural Networks: A Dual Approach
01h 14mRecent Developments in Over-parametrized Neural Networks, Part I
23mEnhancing Neural Network Robustness via Synthesis of Repair Programs
31mARENA: Enhancing Abstract Refinement for Neural Network Verification
31mAbstract Neural Networks
17mInput-Relational Verification of Deep Neural Networks
25mProof transfer for fast certification of multiple approximate neural networks
23mPRIMA: General and Precise Neural Network Certification via Scalable Convex Hull