⚡ Hybrid AI: 100 times more energy efficient!

Artificial intelligence is developing at a frenetic pace, and its energy appetite is growing just as much. In the United States, data centers, true electricity ogres, already absorb a significant part of national production and could see their consumption double soon.

Faced with this ecological and economic tension, a promising hybrid path is emerging. It involves marrying the statistical learning of neural networks with symbolic logical reasoning. This blend, called neuro-symbolic AI, would allow systems to perform tasks more accurately while consuming much less electricity.


This method is not designed for simple chatbots, but is intended primarily for robots that interact with the physical world. By integrating vision and movement, it makes them capable of performing actions like manipulating objects with dexterity, while limiting errors related to imperfect vision or clumsy gestures.

The trials, for example on the famous Tower of Hanoi puzzle, are telling. The neuro-symbolic system achieves a much higher success rate than classical models. Furthermore, it requires significantly less learning time and energy, up to 100 times less than what is required by conventional approaches. The savings are simply enormous.

This energy frugality comes at just the right time, as the need for computing power is exploding. Replacing traditional models, often very resource-intensive, with more sober alternatives could thus relieve electrical grids without sacrificing performance.

The work, available on the arXiv platform, suggests that this hybridization could be applied to other sectors. It would then help reduce the carbon footprint of digital technology while enabling new advances in robotics and automation.

Symbolic reasoning in artificial intelligence

Symbolic reasoning works by applying logical rules and abstract concepts to find solutions, somewhat like a human who follows a plan step by step. Unlike statistical learning, which needs immense volumes of data to make predictions, this method deduces results from general principles, which limits the need for heavy computation.

This approach proves very effective for well-structured tasks, such as action planning or object manipulation, where purely statistical models can make mistakes. By following rules, for example on stability or shape, systems avoid random trials and reach their goal more directly, which translates into significant energy savings.

Within a hybrid neuro-symbolic AI, symbolic reasoning complements neural networks by adding a layer of logic that guides decisions. This not only increases accuracy but also makes the systems' operation more understandable, since their actions stem from explicit rules rather than obscure correlations drawn from data.