Symbolic knowledge injection & extraction for autonomous learning

Artificial intelligence (AI) represents one of humanity’s most transformative technological advancements. In recent years, neuro-symbolic (NeSy) AI has emerged as a promising paradigm for building intelligent systems by integrating symbolic and sub-symbolic approaches. By combining structured reasoning with adaptability and scalability, NeSy AI seeks to overcome the limitations of each paradigm, enabling more robust, interpretable, and reliable systems. Within this framework, this thesis focuses on two complementary subfields of NeSy AI: symbolic knowledge injection (SKI) and symbolic knowledge extraction (SKE). SKI allows structured domain knowledge to be incorporated into sub-symbolic models, guiding learning, improving generalization, and enhancing trustworthiness. Conversely, SKE enables the extraction of symbolic, human-understandable knowledge from such models, addressing the opacity of internal representations and supporting interpretability. Together, SKI and SKE form a synergistic loop: injected knowledge influences learning, while extracted knowledge supports inspection, validation, and refinement, fostering explainable and dependable systems. The thesis investigates the challenges and opportunities of NeSy AI with particular emphasis on SKI and SKE in the engineering of intelligent systems. After reviewing background concepts and the state of the art, we identify key challenges in integrating symbolic and sub-symbolic paradigms. We then present original contributions, including methodologies, algorithms, and tools aimed at advancing NeSy systems. The emergence of large language models (LLMs) has further reshaped the AI landscape, offering powerful capabilities for natural language understanding and generation. This thesis explores how LLMs can enhance SKI and SKE, enabling hybrid systems that learn and adapt autonomously. We demonstrate how such approaches can address complex real-world challenges while preserving interpretability and ethical alignment. Finally, we discuss how these contributions advance the broader NeSy vision and support the overarching goal of this thesis: enabling intelligent systems capable of fully autonomous learning through the integration of symbolic reasoning, sub-symbolic learning, and LLMs.