Skip to content
Meet the Icons of the Gaming World — Master Your Money — The Strip's Premier Gaming Experience — Football

Breakthrough DecoHD Method Slashes AI Memory Use by 97% Without Losing Accuracy

A game-changing leap in hyperdimensional computing could power next-gen wearables and sensors. Imagine AI that's 277x more energy-efficient—without sacrificing performance.

,   and  Administrator
1 min read
The image shows an open book with a variety of machines and text on it. The book is filled with...
The image shows an open book with a variety of machines and text on it. The book is filled with pictures of various machines, each with its own unique design and purpose. The text on the book provides further information about the machines and their functions.

Breakthrough DecoHD Method Slashes AI Memory Use by 97% Without Losing Accuracy

Researchers at the University of California, Irvine have developed a new method to make hyperdimensional computing far more efficient. The technique, called DecoHD, drastically cuts memory use while keeping accuracy high. This breakthrough could help deploy energy-saving machine learning in real-world devices like wearables and smart sensors. Hyperdimensional computing (HDC) is known for its energy efficiency but often struggles with high memory demands. To tackle this, Sanggeon Yun, Hyunwoo Oh, Ryozo Masukawa, and Mohsen Imani created DecoHD, which reduces memory needs without losing performance. Their approach uses a learned decomposition of model parameters, staying true to HDC's core principles.

The team also introduced Class-Selective Decomposition (CSD), which breaks down high-dimensional vectors into smaller, class-specific parts. This makes representations more compact and speeds up computations. As a result, DecoHD can cut trainable parameters by up to 97 percent while matching the accuracy of traditional HDC. When tested on dedicated hardware, DecoHD outperformed existing systems. An ASIC implementation showed up to 277 times better energy efficiency and 35 times faster inference than a CPU. It also surpassed GPU performance and previous HDC ASICs in both speed and power savings. The study highlights its potential for wearable health monitors, IoT sensors, and wireless networks, where low power and real-time processing are critical.

DecoHD delivers major improvements in memory efficiency, speed, and energy use for hyperdimensional computing. The method achieves near-identical accuracy with far fewer parameters, making it practical for edge devices. Its hardware performance—277 times more energy-efficient and 35 times faster than CPUs—opens doors for wider adoption in resource-limited applications.

Read also:

Related

Latest