Scalable Neural Learning for Verifiable Consistency with Temporal Specifications

Formal verification of machine learning models has attracted attention recently, and significant progress has been made on proving simple properties like robustness to small perturbations of the input features. In this context, it has also been observed that folding the verification procedure into training makes it easier to train verifiably robust models. In this paper, we extend the applicability of verified training by extending it to (1) recurrent neural network architectures and (2) complex specifications that go beyond simple adversarial robustness, particularly specifications that capture temporal properties like requiring that a robot periodically visits a charging station or that a language model always produces sentences of bounded length. Experiments show that while models trained using standard training often violate desired specifications, our verified training method produces models that both perform well (in terms of test error or reward) and can be shown to be provably consistent with specifications.