What Is Neurosymbolic AI ? An Overview and Frontier Problems

The success of deep learning since the beginning of the 2010s has simultaneously highlighted its shortcomings when it comes to explainability, trust, efficiency, and generalization. These limitations have reinvigorated research in classical symbolic methods for Artificial Intelligence (AI), as such methods are governed by semantically meaningful and interpretable symbols in explicit structures that facilitate explainability, verification, and other desirable properties. Broadly speaking, neurosymbolic AI seeks to integrate both of these fields of deep learning using neural networks and symbolic reasoning to achieve the best of both worlds in an integrated architecture. Given the advent and proliferation of foundation models like Large Language Models (LLMs), such an integration is imperative to instill trust in these models and facilitate a safer adoption for the same. In this chapter, we provide a high-level overview of neurosymbolic AI and present frontier problems in the field to foster new research directions.