The big knock against artificial intelligence is that it will use up too much energy and wind up being more trouble than it’s worth. Indeed: The International Energy Agency estimated that U.S. AI and data centers used about 415 terrawatt hours of power in 2024, more than 10 percent of that year’s nationwide energy output — and that’s expected to double by 2030.
Researchers in the School of Engineering at Tufts University have developed a proof-of-concept for efficient AI systems that could use 100 times less energy than current ones while at the same time provide more accurate results on tasks.
The approach, developed in the laboratory of Matthias Scheutz, Karol Family Applied Technology Professor, uses neuro-symbolic AI — a combination of conventional neural network AI with symbolic reasoning similar to the way humans break down tasks and concepts into steps and categories.
“By breaking it down into small chunks that are manageable — and for each of these chunks, we learn the low-level behavior — we can save time training,” Scheutz said. “We don’t have to train the whole thing; we only have to train the sub-skills.”
Scheutz was quick to point out that the team focuses its work on robots interacting with humans, so the AI tech they employ is not the type of screen-based large language models (LLMs) that are currently out there. The team instead studies visual-language-action (VLA) models, an extension of LLMs with visual and movement capabilities for robots. These models use camera and language inputs and respond by generating actions in the real world, like moving a robot’s wheels, legs, arms, and fingers.
“LLMs are based on the transformer architecture, while a neuro-symbolic model is typically some kind of neural network structure,” he said.
“And the model in our case has a lower-level neural network structure that learns control policies,” he added. “And up top it has a symbolic representation of the task in terms of a logical language that allows it to explicitly reason about world states. So, it will have information such as “I'm holding something,” or, “The cup is on the table,” those kinds of descriptions. These two layers are tied together and when it learns a task, it co-learns the bottom and the top. The whole point here is that it learns the symbolic description and the controls at the same time.”
"Like an LLM, VLA models act on statistical results from large training sets of similar scenarios, but that can lead to errors," said 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."
Scheutz is optimistic about the future of the technology: The neuro-symbolic system could be trained in just 34 minutes, while the standard VLA model took over a day and a half; training the neuro-symbolic model used only 1 percent of the energy required to train a VLA model; and the energy savings continued during execution of tasks with the neuro-symbolic model using only 5 percent of the energy required for running the VLA.
However, he did curtail his expectations.
“One thing that is important for people to understand is it's not a solution across the board,” Scheutz cautioned. “It's not going to replace transformers and foundation models. I don't want to say I want to replace them, but my point is that there are cases where people now also use these models where they shouldn't be used.”
This article was written by Andrew Corselli, Digital Content Editor, SAE Media Group. For more information, visit here .
