[Paper Review] A Survey on Neural-symbolic Learning Systems

Dongran Yu, Bo Yang+3 more|arXiv (Cornell University)|Nov 10, 2021

Neural Networks and Applications94 references4 citations

TL;DR

This survey proposes a comprehensive taxonomy of neural-symbolic learning systems, integrating neural networks' perception with symbolic systems' reasoning to overcome limitations in AI explainability and generalization. It synthesizes advances in methods, challenges, applications, and future directions, highlighting end-to-end learning, symbolic representation, and knowledge construction as key research frontiers for unified AI systems.

ABSTRACT

In recent years, neural systems have demonstrated highly effective learning ability and superior perception intelligence. However, they have been found to lack effective reasoning and cognitive ability. On the other hand, symbolic systems exhibit exceptional cognitive intelligence but suffer from poor learning capabilities when compared to neural systems. Recognizing the advantages and disadvantages of both methodologies, an ideal solution emerges: combining neural systems and symbolic systems to create neural-symbolic learning systems that possess powerful perception and cognition. The purpose of this paper is to survey the advancements in neural-symbolic learning systems from four distinct perspectives: challenges, methods, applications, and future directions. By doing so, this research aims to propel this emerging field forward, offering researchers a comprehensive and holistic overview. This overview will not only highlight the current state-of-the-art but also identify promising avenues for future research.

Motivation & Objective

• To address the limitations of purely neural or symbolic AI systems by unifying perception and reasoning capabilities.
• To identify and systematize key challenges in neural-symbolic learning, including knowledge representation, rule learning, and symbolic reasoning efficiency.
• To provide a structured taxonomy of neural-symbolic systems based on integration methods, applications, and future research directions.
• To highlight the importance of end-to-end learning, symbolic representation learning, and automatic knowledge construction for scalable AI systems.
• To guide future research by outlining promising directions in explainable AI, robotics, and medical decision-making.

Proposed method

• Proposes a novel three-part taxonomy of neural-symbolic learning systems based on integration methods, applications, and future research directions.
• Classifies methods into different categories such as neuro-symbolic neural networks, differentiable logic, and symbolic integration with deep learning.
• Reviews techniques for symbolic representation learning, including graph-based and heterogeneous representation learning to model complex knowledge.
• Analyzes methods for automatic construction of symbolic knowledge, particularly logic rules from data, with emphasis on Inductive Logic Programming (ILP) and data-driven rule mining.
• Integrates insights from cognitive science, such as dual-process theory (System 1 and System 2), to frame the need for hybrid reasoning and perception.
• Evaluates existing approaches through a structured analysis of their strengths, limitations, and applicability across domains like computer vision, NLP, and healthcare.

Experimental results

Research questions

• RQ1 How can neural-symbolic systems effectively combine deep learning’s perception with symbolic reasoning’s interpretability?
• RQ2 What are the core technical challenges in enabling end-to-end learning of neural-symbolic models, especially in rule and knowledge acquisition?
• RQ3 How can symbolic representation learning be improved to capture semantic similarities between predicates like 'near' and 'next to'?
• RQ4 What are the most promising application domains for neural-symbolic systems beyond traditional AI tasks?
• RQ5 What future research directions are most critical for advancing the scalability, generalization, and explainability of neural-symbolic AI?

Key findings

• Neural-symbolic learning systems effectively unify perception and reasoning, enabling AI to perform complex tasks requiring both data-driven learning and logical inference.
• Current symbolic knowledge construction remains largely manual, highlighting a major bottleneck in scalability and a key area for future research in automatic rule learning from data.
• Symbolic representation learning struggles to differentiate semantically similar predicates (e.g., 'near' vs. 'next to'), limiting reasoning efficiency and requiring improved embedding methods.
• Graph representation learning, especially heterogeneous graph embeddings, offers a promising path to model complex, multimodal symbolic knowledge with high expressivity.
• Neural-symbolic systems show strong potential in high-stakes domains such as medical diagnosis and autonomous driving, where interpretability and robustness are critical.
• The integration of dual-process theory (System 1 and System 2) provides a cognitive foundation for designing hybrid AI systems that balance speed and reasoning.