A Survey of Model Learning Techniques for Recurrent Neural Networks.

Ensuring the correctness and reliability of deep neural networks is a challenge. Suitable formal analysis and verification techniques have yet to be developed. One promising approach towards this goal is model learning, which seeks to derive surrogate models of the underlying neural network in a model class that permits sophisticated analysis and verification techniques. This paper surveys several existing model learning approaches that infer finite-state automata and context-free grammars from Recurrent Neural Networks, an essential class of deep neural networks for sequential data. Most of these methods rely on Angluin’s approach for learning finite automata but implement different ways of checking the equivalence of a learned model with the neural network. Our paper presents these distinct techniques in a unified language and discusses their strengths and weaknesses. Furthermore, we survey model learning techniques that follow a novel trend in explainable artificial intelligence and learn models in the form of formal grammars.