Neuro-symbolic AI: A machine that can finally handle both logic and learning

• Author:
• Author name

  Quantumrun Foresight
• April 13, 2023

Machine learning (ML) has always been a promising technology with its unique challenges, but researchers are looking to create a logic-based system that goes beyond big data. Logic-based systems are designed to work with symbolic representations and reasoning, which can provide a more transparent and interpretable way of understanding a system's decision-making process. 

Neuro-symbolic AI context

Neuro-symbolic AI (also called composite AI) combines two branches of artificial intelligence (AI). First is the symbolic AI, which uses symbols to understand relationships and rules (i.e., the color and shape of an object). For symbolic AI to work, the knowledge base must be precise, detailed, and exhaustive. This requirement means that it cannot learn by itself and depends on human expertise to keep updating the knowledge base. 

The other component of neuro-symbolic AI is deep neural networks (deep nets) or deep learning (DL). This tech uses numerous layers of nodes that mimic the human brain's neurons to self-learn to process large datasets. For example, deep nets can go through different images of cats and dogs to accurately identify which is which, and they improve over time. However, what deep nets can't do is process complex relationships. By combining symbolic AI and deep nets, researchers use DL to churn large amounts of data into the knowledge base, after which symbolic AI can infer or identify rules and relationships. This combination allows for more efficient and accurate knowledge discovery and decision-making.

Another area that neuro-symbolic AI addresses are deep net's costly training process. Additionally, deep nets can be sensitive to small input data changes, leading to classification errors. They also struggle with abstract reasoning and answering questions without much training data. Furthermore, the internal workings of these networks are complex and difficult for humans to understand, making it a challenge to interpret the reasoning behind their predictions.

Disruptive impact

Researchers from Stanford University conducted initial studies of composite AI using 100,000 images of basic 3D shapes (squares, spheres, cylinders, etc.) They then used different questions to train the hybrid to process data and infer relationships (e.g., are the cubes red?). They found that neuro-symbolic AI could answer these questions correctly 98.9 percent of the time. Additionally, the hybrid only required 10 percent of training data to develop solutions. 

Since symbols or rules control deep nets, researchers can easily see how they are “learning” and where breakdowns occur. Previously, this has been one of the weaknesses of deep nets, the inability to be tracked because of layers and layers of complex codes and algorithms. Neuro-symbolic AI is being tested in self-driving vehicles to recognize objects on the road and any changes in the environment. It is then trained to react appropriately to these external factors. 

However, there are differing opinions on whether the combination of symbolic AI and deep nets is the best path toward more advanced AI. Some researchers, such as those from Brown University, believe that this hybrid approach may not match the level of abstract reasoning achieved by human minds. The human mind can create symbolic representations of objects and perform various types of reasoning using these symbols, using biological neural networks, without needing a dedicated symbolic component. Some experts argue that alternative methods, such as adding features to deep nets that mimic human abilities, may be more effective in enhancing AI capabilities.

Applications for neuro-symbolic AI

Some applications for neuro-symbolic AI may include:

• Bots, such as chatbots, which can better understand human commands and motivation, producing more accurate responses and services.


• Its application in more complex and sensitive problem-solving scenarios such as medical diagnosis, treatment planning, and drug development. The tech can also be applied to accelerate scientific and technological research for fields such as transportation, energy, and manufacturing. 


• The automation of decision-making processes that currently require human judgment. As a result, such applications may lead to a loss of empathy and accountability in certain fields like customer service.


• More intuitive smart appliances and virtual assistants that can process different scenarios, such as proactively conserving electricity and implementing security measures.


• New ethical and legal questions, such as issues related to privacy, ownership, and responsibility.


• Improved decision-making in government and other political contexts. This technology could also be used to influence public opinion through more targeted advertising and the generation of hyper-personalized advertisements and media.

Questions to consider

• How else do you think neuro-symbolic AI will affect our day-to-day lives?


• How can this technology be used in other industries?