Neuro-symbolic systems: a perspective of uncertainty management

As an intelligent software paradigm, neuro-symbolic systems – integrating neural networks and symbolic rules – have garnered increasing attention, with explosive growth in related research.However, existing approaches still struggle with the effective and efficient construction of large-scale neuro-symbolic systems.One key factor contributing to the unsatisfactory performance of neuro-symbolic learning is the incompatible management of uncertainty during both neural network training and symbolic rule learning processes.This paper aims to provide a systematic, consistent, and rigorous perspective to facilitate the discussion and analysis of the technical challenges in “end-to-end” neuro-symbolic learning.Based on this perspective, the paper organizes and reviews the ideas, progress, and current issues in related work, while also looking ahead to future directions in the development of this field.