AMHERST, Mass. — As artificial intelligence systems grow larger and more powerful, their energy demands are rising dramatically. But recent research from the University of Massachusetts Amherst published Monday in Nature Communications suggests that advanced AI capabilities may be achievable with dramatically lower energy consumption.
A team led by Hava Siegelmann , Provost Professor in the Manning College of Information and Computer Sciences at UMass Amherst, has developed novel AI that more closely mirrors key aspects of how the human brain operates. Siegelmann and her lab have focused on two complementary goals: enabling AI systems to learn continuously in real time rather than only during a fixed training phase, and dramatically reducing the energy required for intelligent computation.
"Current AI systems are extraordinarily powerful, but they are also extraordinarily energy-hungry," said Siegelmann. "Our work shows that it is possible to design AI that remains highly capable while operating much more efficiently."
The contrast with the human brain is striking. According to the National Institutes of Health, the human brain contains approximately 86 billion neurons operating in parallel while consuming roughly 20 watts of power—about the same as a small LED light bulb. By comparison, training the largest AI models can require tens of millions of watts of power and occupy massive data-center infrastructure.
One reason the human brain operates so efficiently is that it functions asynchronously. Only small subsets of the brain's neurons fire when participating in a particular task, or "updating" themselves. This allows complex behavior to emerge with remarkably low energy consumption.
Unlike biological brains, which operate through largely asynchronous activity, today's deep network AI systems, including ChatGPT and Claude, rely on highly synchronized computations across millions or billions of artificial neurons. Updates occur simultaneously, and are governed by a global clock, regardless of the underlying task. Today's networks have to perform substantial computation, requiring intensive energy usage, merely to maintain synchronized operation.
"This synchronized approach worked well when neural networks were relatively small," Siegelmann explained. "But as AI systems have grown to billions and even trillions of parameters, the energy requirements have become increasingly costly, environmentally significant, and impractical for autonomous systems such as robots."
"Researchers have previously explored asynchronous spiking neural networks as a route to energy efficiency," she continued, "but they have found significant obstacles to learning and adaptation because training procedures that work for spiking architectures are far inferior to gradient-based training methods, including backpropagation that have made modern deep neural networks so successful."
Siegelmann's new research team sought to combine the advantages of both approaches.
The result is ANT—Asynchronous Neural Turing networks—a new architecture that removes the need for global synchronization while preserving the differentiable neural properties that make deep neural networks highly trainable.
"The core challenge was eliminating the synchronizing global clock without sacrificing computational power or adaptability," said Siegelmann. "We developed new design principles that allow information to be preserved during asynchronous updates while maintaining powerful learning capabilities."
Because ANT updates only those neurons needed at each computational step, it can reduce energy consumption by orders of magnitude.
Siegelman continued, "In principle, ANT can match the computational power of conventional digital systems and modern deep neural networks while operating efficiently."
The work builds on Siegelmann's longstanding contributions to theoretical neural computation, including her landmark 1995 demonstration that recurrent neural networks possess computational capabilities comparable to Turing machines.
The research team is now working to further improve ANT's energy efficiency and expand its capabilities for continual, real-time learning.
Siegelmann hopes the new framework will encourage broader exploration of AI systems that are not only more sustainable, but also more adaptive and potentially more capable than today's dominant architectures.
Beyond reducing the environmental footprint of AI, the approach may prove particularly valuable for intelligent autonomous systems operating under tight energy constraints, including robots, edge-computing devices, autonomous vehicles and future generations of adaptive machine intelligence.
This research was supported by the U.S. National Science Foundation and the Air Force Office of Scientific Research.
