What Is Neuro-Symbolic AI?

Share This:

Facebook

Twitter

LinkedIn

Email

November 11, 2025

Summary:

• Neuro-symbolic AI blends two major branches of artificial intelligence: neural networks, which learn to identify patterns from data, and symbolic reasoning systems, which apply explicit rules and logic to draw conclusions.
• It’s often referred to as the third wave of artificial intelligence, and has drawn significant attention from the research community and businesses because it can address weaknesses each approach has on its own. 
• It’s already used by some large businesses, such as robots in warehouses, and has promising applications in other autonomous systems. 

Neuro-symbolic AI is often referred to as the “third-wave” of artificial intelligence because it combines two types of AI to solve problems each one struggles with on its own. Its name provides clues as to how it works. 

The “neuro” part of the name refers to neural networks, which have become the dominant AI approach since the 1980s. They are great at learning from data. They recognize patterns and they power the large language models behind generative AI systems. 

The “symbolic” portion of the name refers to symbolic reasoning systems, which were part of a first wave of AI research dating back to the 1950s. It involved teaching computers to follow clear, logical rules and make inferences from them. A symbolic system might have a rule stating, “All mammals have lungs.” It can infer “Whales have lungs” if it knows whales are mammals.

In this Q&A,  IEEE Fellow Houbing Song explains why this emerging paradigm has captured the attention of researchers and industry leaders alike.

What’s behind the need to combine these two distinct AI approaches? 

Neural networks excel in unknown environments because they can learn, but their processes are often opaque and difficult to understand. We don’t know how these systems make the decisions they make, leading to poor explainability. Symbolic systems are effective in familiar environments because they follow established rules. Symbolic systems struggle to generalize in new situations, resulting in a lack of robustness. 

Researchers wanted to find a solution to address the explainability challenge faced by neural networks, particularly deep learning, and the generalizability and robustness challenges faced by symbolic reasoning. 

Was this dual approach of ‘learning’ and ‘reasoning’ inspired by any specific models from psychology or cognitive science?

Neuro-symbolic AI mirrors psychologist Daniel Kahneman’s idea that people think in two ways — one fast and intuitive, the other slow and logical.

The fast and intuitive system is very good at pattern recognition and is a lot like deep learning and neural networks. The slow and logical style of thinking is a lot like symbolic reasoning systems: good at planning, deduction and deliberative thinking.

Both are needed for a robust, assured and trustworthy AI that can learn, reason and interact with humans to accept advice and answer questions. 

How does neuro-symbolic AI improve on today’s mainstream AI systems?

Compared with today’s mainstream AI systems, neuro-symbolic AI systems are able to understand the concepts they reason over and operate with, change their behavior appropriately in response to contextual interventions and align their operations with societal expectations and human intentions.

Can you give a simple real-world example of where neuro-symbolic AI would shine? From a recent lecture you gave for the IEEE Intelligent Systems Council , it appears to have promise in autonomous vehicles. Why does neuro-symbolic AI improve on current tech? 

Autonomous systems, enabled by neuro-symbolic AI, will operate safely and perform as intended. For example, some very large companies are using neuro-symbolic AI to enhance the performance, reliability and safety of their robotics fleet in fulfillment centers. They use symbolic reasoning for spatial problems, such as figuring out the optimal place to pick up an item or put it down, and use neural networks to enable them to handle perception tasks, including categorizing images.

What are the biggest technical challenges in making neuro-symbolic AI work?

The biggest technical challenge in making neuro-symbolic AI work is how to integrate neural networks with symbolic reasoning. Researchers have tried several ways to combine them — running the systems one after another (serial), side-by-side (parallel), placing a neural component inside a symbolic engine or putting a symbolic engine inside a neural network.  However, the best approach for integration has yet to be discovered.

Learn More: If you are interested in a deeper dive into neuro-symbolic systems, check out IEEE Fellow Houbing Herbert Song’s webinar on the topic, presented by the IEEE Systems Council.