Verifying and Interpreting Neural Networks Using Finite Automata
Marco Sälzer, Eric Alsmann, Florian Bruse, Martin Lange
Published in Information and Computation, 308, article 105398, 2026
We translate the input-output behavior of deep neural networks into weak Büchi automata. This representation lets us apply automata-based methods to verification and interpretation questions, including adversarial robustness and minimum sufficient reasons.
The journal article appeared in Information and Computation, volume 308, January 2026. An earlier conference version appeared at DLT 2024, pages 266–281.