KIST develops neuromorphic AI training method that outperforms Google's with far less computing load

As generative AI tools represented by ChatGPT have proliferated, the soaring power consumption of data centers has emerged as a pressing global challenge. To address it, major tech companies around the world are racing to develop low-power "neuromorphic semiconductors" modeled on the human brain.

Spiking neural networks, or SNNs — a key technology in that effort — can dramatically cut power use by transmitting signals only when needed. But they have struggled to match the performance of deep neural networks, or DNNs, which underpin current AI systems such as ChatGPT.

A team led by Dr. Park Sung-sik at the Korea Institute of Science and Technology's Semiconductor Technology Research Division has developed a new training technique called A²SG that significantly boosts SNN learning performance. The findings were accepted as a full paper at ICML 2026, one of the world's three leading AI conferences.

Applied to a large transformer-based SNN — the same core architecture behind ChatGPT — A²SG achieved world-best accuracy among spiking neural networks on the ImageNet large-scale image recognition benchmark. In AI training, a model is refined by solving problems, checking results and adjusting parameters to find better answers; "gradients" guide how much and in which direction each parameter should change. SNNs transmit signals differently from DNNs, making it difficult to tune them properly using the gradient methods developed for DNNs. A²SG overcomes this by combining an "adaptive" approach that adjusts the training method based on learning conditions with an "asymmetric" approach that reflects the characteristics of biological neurons, together yielding stronger training performance.

A²SG delivers world-class performance at a fraction of the cost. It recorded higher accuracy than Google's leading training technique while imposing only about one-sixth the additional computational overhead, and demonstrated consistent performance gains across a range of neural network architectures and applications — from small to large models — confirming its versatility.

Because A²SG can be applied purely in software without any hardware modifications, it is suited for low-power applications such as on-device AI in smartphones, wearables and drones, as well as always-on smart sensors. The research team plans to pursue commercialization of low-power AI semiconductors following large-scale model training and neuromorphic hardware validation, and intends to apply the training algorithm technology to AI model development for next-generation AI chips — including the RPU (Random Processing Unit), a probability-based next-generation semiconductor under development at KIST.

"This research will serve as a stepping stone for neuromorphic AI to close the gap with conventional systems and accelerate the era of next-generation low-power AI," Park said. "We plan to extend the core training algorithm technology we have secured into training techniques for AI models that will run on next-generation AI semiconductors."