[Paper Review] Graph Neural Networks Meet Neural-Symbolic Computing: A Survey and Perspective

Luís C. Lamb, Artur d’Avila Garcez+4 more|arXiv (Cornell University)|Feb 29, 2020

Advanced Graph Neural Networks52 references113 citations

TL;DR

This paper surveys the intersection of Graph Neural Networks (GNNs) and neural-symbolic computing, outlining taxonomy, relationships, and promising directions for integrated reasoning and learning.

ABSTRACT

Neural-symbolic computing has now become the subject of interest of both academic and industry research laboratories. Graph Neural Networks (GNN) have been widely used in relational and symbolic domains, with widespread application of GNNs in combinatorial optimization, constraint satisfaction, relational reasoning and other scientific domains. The need for improved explainability, interpretability and trust of AI systems in general demands principled methodologies, as suggested by neural-symbolic computing. In this paper, we review the state-of-the-art on the use of GNNs as a model of neural-symbolic computing. This includes the application of GNNs in several domains as well as its relationship to current developments in neural-symbolic computing.

Motivation & Objective

• Summarize neural-symbolic computing and its taxonomy.
• Relate GNN models to neural-symbolic computing approaches.
• Survey GNN architectures for relational and symbolic learning.
• Identify challenges and directions for integrating symbolic reasoning with neural learning.

Proposed method

• Present a taxonomy of neural-symbolic systems based on Kautz's framework.
• Describe the historical connection between GNNs and neural-symbolic computing.
• Explain how graph convolutions and attention mechanisms enable relational and symbolic reasoning.
• Discuss Logic Tensor Networks and relational embeddings as NSC tools.
• Explain Pointer Networks and their role for set-based, combinatorial tasks.
• Highlight opportunities for integrating GNNs with NSC towards type 6 symbolic reasoning.

Experimental results

Research questions

• RQ1 How do Graph Neural Networks relate to neural-symbolic computing across different system types?
• RQ2 What are the main NSC architectures that can be expressed through GNNs and related graph-based methods?
• RQ3 What directions or architectures could enable true symbolic reasoning inside a neural engine?
• RQ4 What are the limitations of current NSC approaches when applied to graph-structured data?

Key findings

• GNNs are framed as a natural type of neural-symbolic system within Kautz's taxonomy.
• Graph representations provide permutation invariance and scalable handling of variable graph sizes for reasoning tasks.
• Logic Tensor Nets and relational embeddings illustrate how symbolic knowledge can guide neural learning.
• Attention and pointer-based mechanisms are used to tackle structured, combinatorial problems on graphs.
• The paper outlines research directions toward fully integrated symbolic reasoning within neural architectures (type 6).