Achieving World-Leading Performance in Brain-Inspired Spiking Neural Networks
Securing Core Original Technology for AI Semiconductors... Selected by Top AI Conference ICML

As generative artificial intelligence (AI) becomes more widespread, the massive power consumption of data centers is emerging as a new social issue. Against this backdrop, a team of Korean researchers has developed a groundbreaking core learning technology that can dramatically enhance the performance of next-generation AI operating with low power, much like the human brain. The technology is considered a fundamental advancement that could accelerate the commercialization of low-power AI semiconductors.


The Korea Institute of Science and Technology (KIST) announced on July 23 that a research team led by Senior Researcher Seongsik Park of the Semiconductor Technology Research Division has developed a new learning method, called "A²SG (Adaptive and Asymmetric Surrogate Gradients)," which significantly boosts the training performance of spiking neural networks (SNNs), the core of neuromorphic AI.

Comparison of learning landscapes achieved by AI models according to training methods. Provided by the research team

Comparison of learning landscapes achieved by AI models according to training methods. Provided by the research team

View original image

The research findings have been accepted as a regular paper at ICML 2026, the International Conference on Machine Learning, which is one of the world’s three leading AI conferences.


Spiking neural networks operate by transmitting electrical signals only when necessary, similar to the human brain, which can drastically reduce power consumption compared to conventional deep learning. However, limited learning efficiency has made it difficult for SNNs to achieve deep neural network (DNN)-level performance, such as that seen in generative AI models like ChatGPT.


Improved AI Performance with Reduced Computational Load


The A²SG developed by the research team is characterized by combining an adaptive method, which allows AI to autonomously adjust learning directions during training, with an asymmetric approach that reflects the properties of actual brain neurons. This combination significantly enhances the learning efficiency of existing spiking neural networks.


According to the researchers, when A²SG was applied to a large-scale spiking neural network based on the transformer architecture—the core structure of ChatGPT—the model achieved world-leading accuracy on ImageNet, which is a standard benchmark in image recognition.


Notably, efficiency was improved along with performance. Compared to Google's leading spiking neural network training techniques, A²SG required only about one-sixth the additional computational resources, yet delivered even higher accuracy. Consistent performance improvements were demonstrated across a range of neural network architectures, from small to large models, verifying its versatility.

Dr. Seongsik Park's research team from KIST presenting their research achievements at ICML 2026. (From left) Senior Researcher Seongsik Park (corresponding author), Master's student researcher Yechan Kang (first author), and Master's student researcher Sohee Park (co-author) are taking a commemorative photo at the ICML 2026 poster presentation venue held at COEX in Seoul. Provided by KIST

Dr. Seongsik Park's research team from KIST presenting their research achievements at ICML 2026. (From left) Senior Researcher Seongsik Park (corresponding author), Master's student researcher Yechan Kang (first author), and Master's student researcher Sohee Park (co-author) are taking a commemorative photo at the ICML 2026 poster presentation venue held at COEX in Seoul. Provided by KIST

View original image

This achievement is also significant in that it demonstrates a domestic research team securing original technology in the field of neuromorphic AI training algorithms, an area previously dominated by global big-tech and overseas universities.


The team expects that the technique can be widely applied in on-device AI fields—where low power is critical—such as smartphones, wearable devices, drones, and always-on smart sensors, because it can be implemented purely through software without hardware modifications. Looking ahead, the plan is to link this technology with "RPU (Random Processing Unit)," the next-generation AI semiconductor under development at KIST, in order to promote the commercialization of low-power AI semiconductors.



Seongsik Park, Senior Researcher at KIST, stated, "By solving the structural learning challenges that have hindered the performance of neuromorphic AI, we have increased the practical potential of low-power AI. We aim to contribute to the era of low-power AI by developing AI models that run on next-generation AI semiconductors."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing