Online Learning for Neural Network Verification

Formal verification of neural networks is an essential component for deploying AI in safety-critical applications, but current approaches often struggle with scalability and efficiency. This work investigates how online learning techniques can dynamically improve branching in neural network verification frameworks like Marabou, Beta-Crown, and NeuralSAT. We are developing reinforcement learning methods that adaptively select branching variables during the verification process, potentially reducing verification time. Our preliminary explorations suggest that adaptive, learned heuristics may outperform static branching rules across various network architectures and verification queries. This research aims to address key bottlenecks in neural network verification, contributing to more scalable and reliable safety guarantees for deep learning systems.