A path has opened to run high-performance artificial intelligence (AI), once dependent on large servers, on smartphones.
A team led by Professor Noh Joon-seok of the departments of mechanical engineering, chemical engineering, electrical engineering, and the graduate school of convergence at Pohang University of Science and Technology POSTECH said on the 7th it developed a technology that cuts AI computation by more than 99% while maintaining performance. The study was conducted jointly with researchers at Tsinghua Shenzhen International Graduate School, Harbin Institute of Technology, and City University of Hong Kong, and was published in March in the international journal "Nature Communications."
The team focused on "complex-valued neural networks." While conventional AI mainly handles magnitude information of numbers, complex-valued neural networks can also process phase information that indicates when a signal arrives. Because of this, they have high potential in fields that require precise signal processing, such as holography, wireless communications, and radar image analysis.
However, complex-valued neural networks require heavy computation, making them hard to use on small devices. Quantization, an existing AI lightweighting technique, is mostly designed for conventional neural networks; when applied to complex-valued networks, it distorts phase information and degrades performance.
To solve this, the researchers developed a "joint quantization" method that considers the real and imaginary parts of complex numbers together rather than compressing them separately. They added an adaptive mixed-precision training strategy that keeps important parts at high precision and processes less important parts at lower precision.
In experiments, compared with the previous state-of-the-art model HoloNet in hologram generation, computation was reduced by 99.1% and memory usage by 99.8%. The PSNR metric representing image quality improved by about 4 dB. In voice and wireless signal classification and radar target recognition, computation was cut by more than 85% while maintaining accuracy.
In smartphone run tests, the speed was up to 389 times faster than before. The team expects the technology to be used in fields requiring real-time signal processing, including lightweight Augmented Reality (AR) and Virtual Reality (VR) holograms, Autonomous Driving radar, next-generation communication networks, and portable medical devices.
Professor Noh said, "We confirmed the potential to bring high-performance physics computation AI, once possible only on large servers, into smartphones and small devices," adding, "It will form the basis for lightweight AI use in fields with heavy computational loads, such as electromagnetics, thermodynamics, and quantum physics."
References
Nature Communications (2026), DOI: https://doi.org/10.1038/s41467-026-70319-0