Intelligibilité des réseaux de neurones récurrents par interprétation avec des machines à états finis

Deep Learning has emerged as a dominant computing paradigm in the field of AI, demonstrating remarkable performance across a wide array of tasks. However, its inherent opacity presents significant challenges in terms of interpretability. Conversely, Finite State Machines (FSMs) represent a mature computing paradigm with established formal methods for analysis. This thesis aims to explore intersections and divergences between these two paradigms by investigating their interpretability within the framework of computational complexity theory. Our investigation uncovers a notable disparity in interpretability between RNNs and Weighted Automata (WAs). Specifically, we establish that while the complexity of conducting interpretative analysis on WAs, utilizing metrics such as the SHAP score and polarization score, is tractable, the same cannot be asserted for RNN-ReLUs. This observation highlights the potential of automata-based models as promising interpretation proxies for neural-based models, providing a compelling rationale for their incorporation in interpretative analyses. In a separate context, we examined the complexity associated with quantifying disparities between automata-based and neural-based language models. To address this challenge, we developed novel approximation schemes with PAC guarantees for computing the cross-entropy between Stochastic Weighted Automata and various families of RNNs, including Long Short-Term LSTM/GRUs