Hybridization Beyond Modularity: A Modern Taxonomy for Fusion and Duality of Data-driven and Knowledge-based Artificial Intelligence

Abstract

Combining data-driven and knowledge-based techniques is a key research direction for developing trustworthy artificial intelligence (AI). Modular hybridization concepts in particular have recently received increased attention in scholary taxonomization concepts, such as the so-called Boxology. At the same time, non-modular hybrid approaches based on fusion and duality have continued to evolve and diversify during the last decade. Building on the canonical typology by McGarry et al, we introduce an updated taxonomy for hybridization approaches that combine data-driven and knowledge-based AI in a non-modular fashion. Notably, we distinguish unified and transformational hybridization into further subcategories and indicate current research trends in these areas. While the original work by McGarry and colleagues was targeting solely neural hybrid architectures, our taxonomy is not limited to neurosymbolic techniques but also includes further hybrid approaches, such as fuzzy-based techniques. Overall, this work seeks to stimulate critical discussion about recent developments and future directions in non-modular hybridization.