Add BrightSurf on Google Email

KIST develops neuromorphic AI training technique to usher in the era of low-power AI

08.10.26 | National Research Council of Science & Technology
Apple iPhone 17 Pro

Apple iPhone 17 Pro delivers top performance and advanced cameras for field documentation, data collection, and secure research communications.
[Figure 1] Dr. Seongsik Park 's research team at KIST, which presented its research findings at ICML 2026
(From left) Senior Researcher Seongsik Park (corresponding author), Master's student researcher Kang Yechan (first author), and Master's student researcher Park Sohee (co-author) pose for a commemorative photo at the ICML 2026 poster session held at COEX in Seoul. Credit: Korea Institute of Science and Technology(KIST)

While asking questions to ChatGPT and similar tools to generate images has become part of daily life, behind this convenience lies the massive power consumption of huge data centers. As power issues have emerged as a societal challenge, global big tech companies are moving to develop low-power "neuromorphic semiconductors" that mimic the human brain, and South Korea has also designated "AI semiconductors" as a priority technology within its national strategic technology framework to secure self-reliance. The human brain operates on minimal energy by exchanging signals only when necessary. "Spiking neural networks (SNNs)," modeled after the human brain, consume far less power than conventional AI, but they have struggled to match the performance of deep neural networks (DNNs)-the foundation of current AI systems like ChatGPT.

A research team led by Senior Researcher Seongsik Park of the Semiconductor Technology Research Division at the Korea Institute of Science and Technology (KIST; President Oh Sang-rok) announced that it has developed "A²SG," a new learning technique that enhances the learning performance of SNNs. This research was accepted as a regular paper at ICML 2026, one of the world's top three artificial intelligence conferences, and was presented on Tuesday, July 7, at the conference (July 6-11, COEX, Seoul)-the first time in the conference's history that it was held in South Korea.

The A²SG developed by the research team was applied to a large-scale spiking neural network (SNN) based on the transformer-the core architecture of ChatGPT-and achieved world-leading accuracy for spiking neural networks in the large-scale ImageNet image recognition evaluation. Artificial intelligence refines its models by solving problems, verifying results, and iteratively finding better solutions. In this process, "gradients" serve to indicate the direction and extent to which each model should be adjusted. However, because SNNs exchange signals differently from DNNs, it was difficult to fine-tune the model adequately using only the gradients employed in conventional DNNs. A²SG enhanced training performance by combining an "adaptive" approach-which adjusts the tuning method based on the training context-with an "asymmetric" approach that reflects the characteristics of brain neurons.

A²SG is also noteworthy for delivering world-class performance at a lower cost. It achieved higher accuracy while requiring only about one-sixth the computational overhead of Google's leading training method, and demonstrated consistent performance improvements across a wide range of neural network architectures and applications-from small to large models-thus proving its versatility.

This achievement is significant in that KIST has secured core technology in the field of neuromorphic AI learning algorithms, an area previously led by universities and global tech giants. Since A²SG can be implemented using software alone without any hardware modifications, it can be applied to low-power applications such as on-device AI in smartphones, wearable devices, and drones, as well as smart sensors that operate 24 hours a day. Going forward, the research team plans to commercialize low-power AI semiconductors following large-scale model training and verification of neuromorphic hardware. They also intend to apply the acquired training algorithm technology to the development of AI models for next-generation AI semiconductors, such as the probability-based "RPU (Random Processing Unit)" currently under development at KIST.

Seongsik Park , a senior researcher at KIST, stated, "This research addresses the structural challenges in learning that had hindered the performance improvement of neuromorphic AI, thereby increasing the potential for the practical application of low-power AI." He added, "Going forward, we will develop this core technology into AI models that operate on next-generation AI semiconductors, contributing to the dawn of the low-power AI era." The KIST Post-Silicon Semiconductor Institute, which has made it its mission to develop innovative semiconductors that will transform future computing paradigms, plans to link this achievement with its research on next-generation AI semiconductors to secure core foundational technologies for high-efficiency intelligent semiconductors and strengthen the foundation for technological self-reliance.

###

KIST was established in 1966 as the first government-funded research institute in Korea. KIST now strives to solve national and social challenges and secure growth engines through leading and innovative research. For more information, please visit KIST’s website at https://kist.re.kr/eng/index.do

This research was supported by the Ministry of Science and ICT (Minister Bae Kyung-hoon) of Korea through the KIST Institutional Program (26E0020), the Institute of Information & Communications Technology Planning & Evaluation (IITP) (RS-2025-02218733), and the Sejong Science Fellowship of the National Research Foundation of Korea (NRF-2021R1C1C2010454). The findings of this study are scheduled to be published in the latest issue of the Proceedings of the International Machine Learning Conference (PMLR).

10.48550/arXiv.2606.11236

A²SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks

7-Jul-2026

Additional Media

[Figure 5] Comparison of Learned Feature Discrimination Capabilities
This image (t-SNE) displays the features extracted from images by artificial intelligence on a two-dimensional map, where each point represents a single image and the color indicates the type of object. The model trained using the conventional method (left) shows categories mixed together, making them difficult to distinguish, whereas the model trained with A²SG (right) clearly separates the categories. This demonstrates that A²SG enables stable and accurate training of SNN models. Credit: Korea Institute of Science and Technology(KIST)
[Figure 2] Presenting research at the ICML 2026 poster session
Dr. Seongsik Park 's research team at KIST is explaining the details of the A2SG study to researchers at the ICML poster session held at COEX in Seoul. Credit: Korea Institute of Science and Technology(KIST)
[Figure 3] Comparison of the Learning Landscapes Reached by AI Models Based on Learning Methods
Conventional artificial intelligence (DNN) has a gentle learning landscape (left). The middle and right figures show the same learning landscape for the same spiking neural network (SNN), where the search path varies depending on the compass (learning method) used. While the conventional method reaches a steep and rugged point (middle), A²SG reaches a flat point, resulting in improved performance (right). Each figure depicts the terrain surrounding the point reached during learning. Credit: Korea Institute of Science and Technology(KIST)
[Figure 4] Overview of the A²SG Method
(a) Valid intervals of the surrogate gradient and the asymmetric (ASY) function form (b) How the spatiotemporal adaptive surrogate gradient operates by detecting changes in the learning landscape in real time and self-correcting accordingly (c) A schematic diagram showing the vicinity of the point reached by the training process before and after applying A²SG on the loss landscape Credit: Korea Institute of Science and Technology(KIST)

Keywords

Applied Sciences and Engineering

Article Information

DOI
10.48550/arXiv.2606.11236
Article Publication Date
2026-07-07
Article Title
A²SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks

Contact Information

Jungmin Lee
National Research Council of Science & Technology
ljm@nst.re.kr

How to Cite This Article

APA:
National Research Council of Science & Technology. (2026, August 10). KIST develops neuromorphic AI training technique to usher in the era of low-power AI. Brightsurf News. https://www.brightsurf.com/news/LDE0JKG8/kist-develops-neuromorphic-ai-training-technique-to-usher-in-the-era-of-low-power-ai.html
MLA:
"KIST develops neuromorphic AI training technique to usher in the era of low-power AI." Brightsurf News, Aug. 10 2026, https://www.brightsurf.com/news/LDE0JKG8/kist-develops-neuromorphic-ai-training-technique-to-usher-in-the-era-of-low-power-ai.html.