UMass Researchers Develop Energy-Efficient AI Architecture for Power-Constrained Systems
Researchers at the University of Massachusetts Amherst have built a new AI architecture that consumes a fraction of the energy required by today's neural networks while maintaining their learning capabilities. The work, published in Nature Communications, introduces Asynchronous Neural Turing networks (ANT), a brain-inspired approach funded by the U.S. National Science Foundation and the Air Force Office of Scientific Research.
The architecture replaces synchronized neural processing with asynchronous updates, potentially reducing energy consumption by orders of magnitude compared with conventional networks. This matters for developers building systems with tight power budgets: robots, autonomous vehicles, edge-computing devices, and other intelligent machines that operate outside traditional data centers.
How ANT Works Differently
Modern AI systems rely on synchronized computation, where large numbers of artificial neurons update simultaneously under control of a global clock. This approach powers today's deep learning advances but demands substantial computing resources and electricity.
The human brain operates differently. Biological neural systems function asynchronously-only small groups of neurons activate when needed for a specific task. ANT mirrors this design principle. Because only neurons required for a particular computation update, the system conserves energy dramatically.
Hava Siegelmann, who led the research at Manning College of Information and Computer Sciences, said the core challenge was eliminating the global synchronizing clock without sacrificing computational power or adaptability. "We developed new design principles that allow information to be preserved during asynchronous updates while maintaining powerful learning capabilities," she said.
The Energy Efficiency Problem
AI's power demands have drawn scrutiny as training and operating large models consumes significant electricity. Data centers running advanced models face rising costs and sustainability concerns. For battery-powered systems, high energy demands make deployment impractical.
Previous attempts at energy-efficient neural architectures-particularly spiking neural networks-struggled to match the learning performance of modern deep networks. ANT solves this by preserving information during asynchronous updates while retaining strong learning and adaptation capabilities.
What Comes Next
The team plans to improve ANT's efficiency further and expand its ability to support continual learning. Siegelmann hopes the framework will encourage development of AI systems that are more efficient, adaptable, and capable than many current architectures.
For IT and development teams, this work addresses a practical constraint: how to deploy advanced AI capabilities on devices with limited power. The approach may prove particularly valuable for autonomous systems and edge-computing applications where traditional cloud infrastructure isn't viable.
The research builds on Siegelmann's earlier theoretical work demonstrating that recurrent neural networks can achieve computational capabilities comparable to Turing machines.
Your membership also unlocks:
