Formal verification of neural networks?

Machine learning is a popular tool for building state of the art software systems. It is more and more used also in safety critical areas. This demands for verification techniques ensuring the safety and security of machine learning based solutions. However, we argue that the popularity of machine learning comes from the fact that no formal specification exists which renders traditional verification inappropriate. Instead, validation is typically demanded, but formalization of so far informal requirements is necessary to give formal evidence that the right system is build. Moreover, we present a recent technique that allows to check certain properties for an underlying recurrent neural network and which may be uses as a tool to identify whether the system learned is right.