Neuro-Symbolic AI: the new frontier of artificial intelligence

Artificial intelligence has lived through several eras: the symbolic one, based on logical rules, then the statistical-connectionist one, dominated by neural networks and by Large Language Models (LLMs) such as ChatGPT. Today a new paradigm is establishing itself, one that joins the best of both worlds: neuro-symbolic AI.

What is Neuro-Symbolic AI?

Neuro-symbolic AI (NSAI) is a hybrid approach that combines two profoundly different models of artificial intelligence:

• Symbolic AI, which rests on logical rules, symbols and deductive reasoning.
• Neural AI, which uses deep neural networks to learn from data without structured prior knowledge.

This fusion makes it possible to build systems that not only learn from data but also reason in a structured way, close to human reasoning.

Figure: the split between symbolic and neural AI, from the source article.

Why is a hybrid approach needed?

Although LLMs have achieved extraordinary results in linguistics, computer vision and content generation, they show several limits:

• They do not truly understand meaning: they work statistically and predictively, but they do not know why an answer is correct.
• They are opaque: it is hard to explain why an LLM produced a given answer.
• They need enormous quantities of data to work well.
• They can make trivial or dangerous mistakes in critical contexts such as medicine or law.

Neuro-symbolic AI seeks to overcome these limits by combining the flexibility of neural networks with the transparency and interpretability of symbolic rules.

How does a neuro-symbolic system work?

An NSAI system can be seen as a “marriage” between two components:

• The symbolic part handles explicit knowledge, logic and relations between concepts. For example, it knows that “all doctors are human” and “all humans are mortal”, and can therefore deduce that “doctors are mortal”.
• The neural part analyses complex data (text, images, audio) to extract information and recognise patterns.

The integration happens through specific architectures, such as Logic Tensor Networks or Neuro-Symbolic Concept Learner models, which allow neural networks to respect logical constraints and to reason on symbolic bases.

Applications of Neuro-Symbolic AI

Neuro-symbolic AI is particularly promising in fields where understanding and explanation are essential:

• Marketing: analysing and understanding customer attitudes (for example hdtwin.ai).
• Public administration: disseminating complex data (for example budget.g0v.it).
• Finance: credit-risk analysis, where it helps to combine regulatory experience with the customer’s historical behaviour (for example Mopso.eu).

Other emerging fields include robotics, compliance, industrial control and human-machine interaction.

Why is it better than LLMs alone?

In short, neuro-symbolic AI offers a series of advantages over the isolated use of LLMs:

• Greater interpretability: thanks to symbolic rules, the decision process can be explained.
• Greater reliability: it reduces the risk of the “hallucinations” typical of LLMs.
• Data efficiency: it can also work with reduced datasets, exploiting explicit knowledge.
• Causal and abstract reasoning: it goes beyond statistical correlation and can formulate deductive hypotheses.

A look at the future

Neuro-symbolic AI is still evolving, but it is one of the most promising directions for building artificial-intelligence systems closer to human intelligence: systems able not only to “answer” but also to “understand” and “explain”. It is a crucial step towards transparent, reliable and responsible systems, able to earn their users’ trust and to operate in sensitive and complex fields.

Further reading

• Dingli, A., & Farrugia, D. (2023). Neuro-Symbolic AI: Design Transparent and Trustworthy Systems That Understand the World as You Do. Packt Publishing.