Neuro-symbolic AI breakthrough cuts energy consumption by 100x

Artificial intelligence is responsible for worsening the energy crisis. Specifically in the United, energy consumption driven by AI’s growing needs and boom in data centers has reached a critical level.

For instance, according to the International Energy Agency, AI systems and data centers used about 415 terawatt hours of power in 2024.

In a recent breakthrough, the researchers from the Tufts University School of Engineering have developed a more energy-efficient approach, aiming to slash increasing energy needs.

By pivoting from purely data-driven models to a hybrid "neuro-symbolic" approach, the team has demonstrated that AI can become both smarter and vastly more energy-efficient.

In a recent study led by Professor Mtthias Scheutz, the team has developed a neuro-symbolic AI system. What makes this system distinguished from traditional models is the combination of the system marked by neural networks along with symbolic reasoning.

Currently, AI-based tasks can consume up to 100 times more energy than traditional search methods. The recently-developed hybrid model is known for achieving 100 times more energy efficiency while increasing accuracy in performance.

When applied to robotics, this hybrid model, known as visual-language-action (VLA), outperformed standard AI in complex tasks while requiring a fraction of the time and energy to operate.

According to Scheutz, "A neuro-symbolic VLA can apply rules that limit the amount of trial and error during learning and get to a solution much faster. Not only does it complete the task much faster, but the time spent on training the system is significantly reduced."

Drastic resource savings

The efficiency gains of the neuro-symbolic approach are massive. For instance, the training time was reduced from 36+ hours( taken by traditional models) to 34 minutes based on this approach.

Energy consumption was also slashed dramatically. For example, the neuro-symbolic model was trained by using only 1 percent of energy of standard models. During operation, it used only 5 percent energy.

Superior performance results

The researchers used the Tower of Hanoi puzzle to compare the systems. The neuro-symbolic system accomplished a 95 percent success rate. On the other hand, traditional models managed to achieve only 34 percent success.

When encountering the complex versions of the puzzle, the hybrid model succeeded 78 percent of the time, whereas the traditional models failed drastically.

A sustainable way forward

According to the researchers, the neuro-symbolic AI offers a more sustainable future by laying down a more dependable and low-energy foundation for AI systems.

Moreover, the system can also fix the problem of hallucinations and logical errors.