N'Dah Jean Kouagou successfully completed her doctorate on the topic of "Fast Neuro-Symbolic Approaches for Class Expression Learning" under the supervision of Prof. Dr. Axel-Cyrille Ngonga Ngomo.
Summary of the doctoral dissertation:
With the swift progress of current AI systems' capabilities, there is a pressing need for explainable AI methods to understand the behavior of these systems. To this end, class expression learning in description logics holds enormous potential. A class expression provides a clear description of why a given example is classified as positive, and is hence a white-box model. Most existing approaches for class expression learning are search-based methods which generate many candidate class expressions and select the one with the highest classification score. While these approaches often perform well on small datasets, their search space is infinite and its exploration becomes arduous on large datasets even for a single learning problem. In this thesis, we develop several neuro-symbolic approaches to tackle class expression learning at scale. Our first approach CLIP uses neural networks to learn concept lengths, and utilizes trained concept length predictors to solve learning problems efficiently. Neural class expression synthesizers (NCES) tackle class expression learning in $\mathcal{ALC}$ in a fashion akin to machine translation, and support sophisticated computing hardware such as GPUs. NCES2 extends NCES to the description logic $\mathcal{ALCHIQ}^{(\mathcal{D})}$, integrates an embedding model for end-to-end training, and employs a data augmentation technique to enhance generalization to unseen learning problems. Finally, ROCES uses iterative sampling to improve the robustness of neural class expression synthesizers to changes in the number of input examples. Experimental results on benchmark datasets suggest that our proposed approaches are significantly faster than the state of the art (up to 10,000$\times$ with NCES, NCES2 and ROCES) while being highly competitive in predictive performance.
