Lookahead Lemma Framework Increases Neural Network Verification Success by 34%

An inprocessing framework for neural network verification uses lookahead procedures to derive lemmas over unstable ReLU phases. Integrated into Marabou and $\alpha$-$\beta$-CROWN, the method utilizes implication graphs to prune search spaces, proving up to 34% more instances unsatisfiable.