A Survey on Neural-symbolic Systems

Dongran Yu, Bo Yang, Dayou Liu, Hui Wang

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. This paper surveys the advancements in neural-symbolic learning systems from four distinct perspectives: challenges, methods, applications, and future directions, offering a comprehensive overview of the current state-of-the-art and promising avenues for future research.

The Evolution of AI Paradigms

For decades, artificial intelligence research has oscillated between two primary doctrines. Symbolic systems (the "First Wave") are excellent at logical reasoning and are inherently interpretable, yet they struggle with unstructured data like images and video. Conversely, Neural networks (the "Second Wave") excel at perception and learning from raw data but operate as "black boxes," lacking transparent reasoning and requiring massive datasets to generalize.

The historical progression of AI from logic-based reasoning to data-driven machine learning.

The challenge addressed by Neural-symbolic (NeSy) systems is the unification of these two paradigms into a single framework—often described as the transformation from System 1 (intuitive/fast) to System 2 (logical/slow) intelligence. Neither approach alone is sufficient for human-like intelligence; neural networks often fail in scenarios requiring strict logic, while symbolic systems cannot handle the messiness of the real world.

Taxonomy of Neural-Symbolic Integration

This survey proposes a novel taxonomy for NeSy systems based on how neural and symbolic components interact:

  1. Learning for Reasoning (Serialization): Neural networks act as accelerators or transformers. They extract symbols from raw data (e.g., identifying objects in a scene) to feed into a symbolic engine or prune the search space of logical inference.
  2. Reasoning for Learning (Parallelization): Symbolic knowledge guides or constrains the training of a neural network. Logical rules are often converted into regularization terms in a loss function, effectively "teaching" the network the laws of the domain.
  3. Learning-Reasoning (Interaction): This is a tightly coupled, bidirectional interaction. A neural network proposes hypotheses, while a symbolic engine evaluates logical consistency, creating an iterative feedback loop to refine the model end-to-end.

Key Insights and Advantages

NeSy systems move away from opaque "black-box" models toward "gray-box" architectures. By leveraging symbolic knowledge, these systems provide a trace of their reasoning, which is vital for Explainable AI (XAI). Furthermore, because symbolic rules provide strong priors, NeSy models often require significantly less data to reach high performance compared to pure deep learning models.

Comparison of Neural, Symbolic, and Neural-symbolic systems across efficiency, generalization, and interpretability.

These hybrid models achieve a "sweet spot" of high efficiency, robust generalization, and interpretability. In complex reasoning tasks like the CLEVR benchmark, NeSy models frequently outperform pure neural architectures because they do not have to "re-learn" the basic laws of logic from scratch.

Future Directions and Challenges

Despite significant progress, several hurdles remain for the field:

  • Intractable Inference: Grounding logic rules in large-scale data can lead to an exponential explosion of possibilities, making real-time reasoning difficult.
  • The Knowledge Bottleneck: Most current systems rely on human-authored logical rules. A major future direction is Automatic Rule Induction, where AI learns to discover underlying logical laws on its own.
  • Semantic Nuance: Improving Symbolic Representation Learning is a priority, as current symbolic logic often fails to capture the semantic overlap between similar concepts (e.g., "near" vs. "next to").

Summary updated

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.
BibTeX
@article{yu2023survey,
  author = {Dongran Yu and Bo Yang and Dayou Liu and Hui Wang},
  title = {{A Survey on Neural-symbolic Systems}},
  journal = {Neural Networks},
  year = {2023},
  doi = {10.1016/j.neunet.2023.06.028},
}
Figure