Compositional Verification without Compositional Specification for Learning-Based Systems

We consider the problem of performing compositional verification of a system with machine learning components whose behavior cannot easily be formally specified. We present an approach involving a system-level verifier communicating with a component-level analyzer wherein the former identifies a subset of environment behaviors that might lead to a system-level failure while the latter identifies erroneous behaviors, such as misclassifications, of the machine learning component that might be extended to a system-level counterexample. Results on cyber-physical systems with deep learning components used for perception demonstrate the promise of this approach.