[LUM#22] Symbolic AI, on the other hand, is based on human reasoning
Logical and efficient, symbolic AI has not yet had its final say in the face of machine learning, which is now gaining ground in the field of computer science. Researchers at the Lirmm are betting on the hybridization of these two types of AI.
In the 20th century, AI did not need to identify itself as symbolic. Originating in the 1950s, methods based on high-level abstract representations—known as “logical” methods—dominated AI research. But over the past decade, with the development of computer processing power, vast amounts of available data, and new algorithms, neural AI—referring to neural networks—has taken the lead. Starting in 2012, traditional AI had to distinguish itself from this new form of AI: it would be called symbolic AI, because it relies on reasoning that uses symbols. Marie-Laure Mugnier, computer scientist at Lirmm1, explains the dominant paradigm of this AI: “For a long time, work on AI was based on the premise that to be intelligent, one must be able to reason. Researchers rely on the representation of human knowledge in mathematical languages that allow for the automation of reasoning. ” The researcher uses the example of a simple deductive argument: Socrates is a man, men are mortal, therefore Socrates is mortal. These three statements are linked by a logical chain that the machine can reproduce.
The first major successes of this AI were in the medical field, where it was used to analyze medical knowledge bases and quickly extract diagnostic information. “The MYCIN expert system was a pioneer; drawing on a knowledge base of approximately 600 rules modeling a physician’s expertise, this program could identify the bacteria responsible for blood infections, such as meningitis, and recommend antibiotic treatments. However, early expert systems were fairly empirical, whereas today’s knowledge-based systems are rigorously grounded in mathematical theories such as logic and probability.” The logical approach is very different from that of deep learning, which relies on complex numerical calculations based on a massive amount of data. This is a major difference, since symbolic AI is, by nature, explainable and therefore likely to be understandable to users.
Analogy with the human brain
“Agronomists fromINRAE came to see us at LIRMM precisely to put an end to ‘black boxes.’ So we worked on a project with the ABSys laboratory to develop an AI system capable of helping them design new agroecological systems,” explains Marie-Laure Mugnier. By leveraging plant databases built by ecologists and formalizing scientific knowledge about the relationships between plant functional traits and ecosystem services into logical rules, the tool can identify species capable of providing certain ecosystem services. A crucial feature of this tool is that it can justify its results. “In viticulture, we specifically tested the identification of plants capable of fixing nitrogen, improving soil structure, or storing water, which would be suitable for vine cover cropping (see: Integrates data and knowledge to support the
selection of service plant species in agroecology, in Computers and Electronics in Agriculture, 2024).”
Symbolic AI remains highly effective today in many fields, such as solving problems modeled as constraint systems. “It could be solving a Sudoku puzzle, but also optimizing an automotive assembly line, ” the researcher points out. To explain the difference between the two types of AI, she offers an analogy by comparing them to the two systems that make up the human brain, according to Daniel Kahneman, psychologist and 2002 Nobel Prize winner in Economics: System 1, which is fast, unconscious, and intuitive, and is used for pattern recognition, is neural AI, while System 2, which is slower, conscious, and explicit, and is used for deduction, is symbolic AI.
Complex, high-level tasks
Neural AI has revolutionized the approach to computing in fields such as image and speech recognition, language translation, and text generation. “But symbolic AI enables the performance of complex, high-level tasks, which remain essential for decision support, planning, or collective deliberation,” notes Marie-Laure Mugnier, who adds in passing that the current craze for machine learning applied to everything is showing its limitations when the volume of data on which the algorithms run is insufficient: “I see many students who, during corporate internships, develop neural AI tools but on datasets that are too small. And it doesn’t work well.” This logical approach, which relies on limited data, offers another advantage to symbolic AI: its frugality. For it is the exponential processing of ever-growing masses of data that is responsible for AI’s ecological footprint.
At Lirmm, several research projects focus on hybrid AI designed to combine the two approaches. “To take the example of agroecological systems, we could identify the plants growing naturally on a plot using image recognition from the Pl@ntNet app, and then use our symbolic AI tool to determine their potential in terms of ecosystem services, ” explains Marie-Laure Mugnier.
UM podcasts are now available on your favorite platform (Spotify, Deezer, Apple Podcasts, Amazon Music, etc.).
- Lirmm (UM, CNRS, Inria, UPVD, UPVM)
↩︎