From Statistical Relational to Neural Symbolic Artificial Intelligence: a Survey

This survey explores the integration of learning and reasoning in two different fields of artificial intelligence: neural-symbolic computation (NeSy) and statistical relational artificial intelligence (StarAI). NeSy aims to integrate symbolic reasoning and neural networks while StarAI focuses on integrating logic with probabilistic graphical models. The survey brings attention to seven shared dimensions between the two approaches. These dimensions are employed to categorize both fields and include: (1) the approach to logic inference, whether model or proof-based; (2) the syntax of logical theories; (3) the logic semantics of the systems and their extensions to facilitate learning; (4) the scope of learning, encompassing either the parameters alone or the entire logic theory; (5) the presence of symbolic and subsymbolic components in representations; (6) the degree to which the systems can capture the original logic, probabilistic, and neural paradigms; and (7) the classes of tasks the systems are applied to. By positioning various NeSy and StarAI systems along these dimensions and pointing out analogies between them, this survey contributes to establishing a common set of fundamental underlying concepts for the integration of learning and reasoning.