Property Checking with Interpretable Error Characterization for Recurrent Neural Networks

This paper presents a novel on-the-fly, black-box, property-checking through learning approach as means for verifying requirements of recurrent neural networks (RNN) in the context sequence classification. Our technique steps on tool probably approximately correct (PAC) deterministic finite automata (DFA). The classifier inside black-box consists Boolean combination several components, including RNN under analysis together with to be checked, possibly modeled themselves. On one hand, if output algorithm is an empty DFA, there proven upper bound (as function parameters) probability language nonempty. implies property holds probabilistic guarantees. other, DFA nonempty, it certain that entails does not satisfy requirement sure. In this case, automaton serves explicit and interpretable characterization error. rely specific specification formalism capable handling nonregular languages well. Besides, neither explicitly builds individual representations any components nor resorts external decision procedure verification. also improves previous theoretical results regarding guarantees underlying algorithm.