Abstract Interpretation of Fixpoint Iterators with Applications to Neural Networks
Explore a groundbreaking abstract interpretation framework for precise over-approximation of numerical fixpoint iterators in this 15-minute video presentation from PLDI 2023. Discover how researchers from ETH Zurich developed CRAFT, a tool that significantly outperforms state-of-the-art methods in speed, scalability, and precision when verifying challenging neural network architectures. Learn about the key theoretical insights and the novel CH-Zonotope abstract domain that enable efficient computation of sound and precise fixpoint abstractions without using joins. Gain valuable knowledge on the application of this framework to monDEQ, a fixpoint-based neural network architecture, and understand its implications for adversarial robustness and equilibrium models in machine learning.