Towards Logical Representations of Recurrent Neural Networks

Representations of neural networks in the formal language of logical systems may be used to enhance their interpretability and as a step in the formal verification of their properties. This work builds on the literature regarding the representation of feedforward neural networks that compute rational McNaughton functions in the language of Łukasiewicz logic. Thus, we propose a technique for representing recurrent computations, enabling the representation of recurrent neural networks. This is an initial investigation conducted within the specific context of simple recurrent networks (SRNs) whose layers compute rational McNaughton functions.