Insider Brief
- Researchers at the University of Massachusetts Amherst have developed a new AI architecture called Asynchronous Neural Turing networks (ANT) that could significantly reduce the energy consumption of advanced AI systems while preserving their learning capabilities, according to a study published in Nature Communications.
- Funded by the U.S. National Science Foundation and the Air Force Office of Scientific Research, the research introduces a brain-inspired approach that replaces synchronized neural processing with asynchronous updates, potentially reducing energy use by orders of magnitude compared with conventional neural networks.
- The researchers said the architecture could be particularly valuable for robots, autonomous vehicles and other power-constrained systems, and are continuing work to improve ANT’s efficiency and ability to support real-time, continual learning.
Researchers at the University of Massachusetts Amherst have developed a new artificial intelligence architecture that could significantly reduce the energy demands of advanced AI systems while preserving their learning capabilities. The work, funded by the U.S. National Science Foundation and the Air Force Office of Scientific Research, outlines a brain-inspired approach that may offer a more efficient path for future AI and robotics applications.
The study, published in Nature Communications, introduces what the team, led by Manning College of Information and Computer Sciences professor Hava Siegelmann, calls Asynchronous Neural Turing networks, or ANT, a new AI architecture designed to more closely resemble how the human brain processes information.
“Current AI systems are extraordinarily powerful, but they are also extraordinarily energy-hungry,” Siegelmann noted. “Our work shows that it is possible to design AI that remains highly capable while operating much more efficiently.”
Addressing the AI Energy Demand
The team’s work comes as the energy demands of training and operating artificial intelligence systems has drawn increased scrutiny. Modern AI models can require enormous computing infrastructure and consume significant amounts of electricity, something that is raising concerns about cost and sustainability. There is also the issue of practicality when it comes to high energy demand with advanced AI when it comes to battery powered systems such as robots and autonomous vehicles.
The study proposes a way to combine the energy efficiency associated with asynchronous computation with the powerful learning capabilities that have made modern neural networks successful.
“The core challenge was eliminating the synchronizing global clock without sacrificing computational power or adaptability,” added Siegelmann. “We developed new design principles that allow information to be preserved during asynchronous updates while maintaining powerful learning capabilities.”
Most modern AI systems rely on synchronized computation, in which large numbers of artificial neurons update simultaneously under the control of a global clock. This approach has helped drive advances in deep learning, but it also requires substantial computing resources and energy.
The researchers note the human brain operates differently. Biological neural systems function asynchronously, with only small groups of neurons becoming active when needed for a particular task. This allows the brain to perform complex computations while consuming relatively little power.
Because only the neurons required for a particular computation are updated, ANT can potentially reduce energy consumption by orders of magnitude compared with conventional neural network architectures, the researchers reported.
To develop ANT, the researchers focused on solving a problem that has challenged previous attempts at energy-efficient neural architectures. While asynchronous spiking neural networks have long been viewed as a promising alternative to conventional AI systems, they have generally struggled to match the learning performance of modern deep neural networks. ANT was designed to preserve information during asynchronous updates while retaining strong learning and adaptation capabilities.
What’s Next?
The researchers indicated they will focus on improving the architecture’s energy efficiency and expanding its ability to support continual learning. Siegelmann said she hopes the framework will encourage the development of AI systems that are more energy efficient, adaptable and capable than many of today’s leading architectures.
The findings could have implications beyond reducing electricity use in data centers. The researchers suggest the approach may be particularly valuable for autonomous systems that must operate with limited power resources, including robots, edge-computing devices, autonomous vehicles and other intelligent machines deployed outside traditional computing environments.
The work builds on Siegelmann’s earlier theoretical research into neural computation, including work demonstrating that recurrent neural networks can achieve computational capabilities comparable to those of Turing machines.
Featured image: Hava Siegelmann, Provost Professor in the Manning College of Information and Computer Sciences at UMass Amherst. (Credit: UMass Amherst)
