New AI Models Could Slash Energy Use While Dramatically Improving Performance

Researchers at the School of Engineering have developed a proof of concept for AI systems that could use 100 times less energy than current ones while providing more accurate results on tasks. The approach combines neural network pattern recognition with symbolic reasoning that breaks down tasks into steps and categories. The work comes from the laboratory of Matthias Scheutz, Karol Family Applied Technology Professor. It focuses on visual-language-action models for robots that use camera and language inputs to generate physical actions such as moving wheels, legs, arms, and fingers. The research will be presented at the International Conference of Robotics and Automation in Vienna in June and published in the conference proceedings. In tests with a standard Tower of Hanoi puzzle, the neuro-symbolic system achieved a 95 percent success rate compared with 34 percent for conventional models. For a more complex version not seen in training, the neuro-symbolic system reached 78 percent success while standard models failed every attempt. Training the neuro-symbolic model took 34 minutes versus over a day and a half for a standard model, using 1 percent of the energy, and task execution used 5 percent of the energy. Scheutz said neuro-symbolic systems apply rules that limit trial and error during learning and reach solutions faster, with significantly reduced training time. He noted that large language models predict the next word or action from statistical patterns, which can produce errors and consume disproportionate energy relative to the task. The researchers conclude that hybrid neuro-symbolic AI could offer a more sustainable and dependable foundation than current large language and visual-language-action models.