Researchers at Tufts University have unveiled a new AI architecture that combines neural networks with human-like symbolic reasoning, potentially cutting energy consumption by 100 times. The breakthrough offers a solution to the surging power demands of modern artificial intelligence, which currently accounts for more than 10% of total electricity production in the United States.
Traditional AI models rely on brute-force statistical prediction, a method that is both energy-intensive and prone to error. The new system, developed in the laboratory of Matthias Scheutz, a professor of applied technology at the university, integrates logical rules to guide decision-making. This approach mirrors human problem-solving by breaking complex tasks into manageable steps.
A hybrid approach to robotics
The team focused their research on visual-language-action (VLA) models, which are used to control physical robots. Unlike standard large language models, these systems must interpret visual data from cameras and translate that information into physical movements, such as operating a robotic hand or arm.
Conventional VLA models often struggle with simple tasks. They rely on massive datasets and trial-and-error learning to understand objects, which frequently leads to mechanical failures or misinterpretations of the environment. Scheutz argues that the current "predictive" nature of these systems is inherently inefficient and prone to inaccuracies.
"A neuro-symbolic VLA can apply rules that limit the amount of trial and error during learning and get to a solution much faster," Scheutz said. "Not only does it complete the task much faster, but the time spent on training the system is significantly reduced."
In performance tests using the Tower of Hanoi puzzle, the hybrid system achieved a 95% success rate, compared to just 34% for standard models. When faced with novel, more complex variations of the puzzle, the neuro-symbolic system succeeded 78% of the time, while traditional models failed every attempt.
The efficiency gains extend to the computational resources required for development. Training the new model required only 1% of the energy consumed by a standard VLA system. During active operation, the model used just 5% of the power typically required by conventional AI architectures.
The research team will present their findings at the International Conference of Robotics and Automation in Vienna this May. By reducing the reliance on massive, energy-draining training sets, the researchers believe their work provides a more sustainable path for the future of robotics.
