Нейроморфный ИИ из KIST обучается в шесть раз быстрее Google — без лишних аппаратных затрат

The South Korean Institute of Science and Technology KIST presented on July 6, 2026 at the ICML 2026 conference the A²SG method (Adaptive and Asymmetric Surrogate Gradients) — a training technique for spiking neural networks that requires approximately six times fewer computational resources than Google's leading method, while achieving comparable accuracy.

What is A²SG and Why It Matters

Spiking neural networks (SNN) mimic the workings of the biological brain: they transmit signals as sparse impulses rather than a continuous stream of numbers, which results in fundamentally lower energy consumption compared to conventional deep neural networks (DNN). The problem is that they are harder to train, since the standard backpropagation method does not work directly for SNNs.

The A²SG method, developed by a team led by Senior Researcher Songsik Park from the KIST Semiconductor Technology Division, addresses this challenge through adaptive and asymmetric surrogate gradients for transformer-based SNN architectures.

• Accuracy gain — world-class level on the ImageNet benchmark
• Computational overhead — ~1/6 of Google's method
• Implementation — software-only, no hardware changes required
• Applicability — smartphones, wearable devices, drones, 24-hour sensors
• Publication — arXiv (DOI: 10.48550/arxiv.2606.11236), ICML 2026, Seoul, July 6–11, 2026

What Changed in the Training Approach

The key difference of this method from its predecessors is that it is simultaneously adaptive and asymmetric. Traditional surrogate gradients apply the same function to all neurons and in all training directions. A²SG adjusts the gradient shape for each neuron depending on its activity and applies different strategies for the forward and backward passes. This is precisely what yields the efficiency gain without sacrificing quality.

"This research eliminates structural problems of neuromorphic AI, enhancing the potential for practical applications," —

Songsik Park, lead researcher at KIST.

According to arXiv, the method showed consistent improvements across neural networks of various sizes and in different tasks. The first author of the paper was graduate student Kang Yechan.

What This Means

Neuromorphic AI is moving out of the academic niche: if A²SG can be commercialized — and KIST has already announced plans to create low-energy AI chips based on RPU (Random Processing Unit) — AI-enabled devices will be able to operate without constant cloud support and with fractions of a watt of consumption instead of the megawatts of data centers.

Frequently Asked Questions

What is a spiking neural network and how is it better than a conventional one?

A spiking neural network (SNN) transmits information as sparse impulses, like brain neurons, rather than as a continuous stream of numbers. This reduces energy consumption by orders of magnitude compared to classical DNNs — while achieving comparable quality results on a number of tasks.

Do I need to change hardware to use A²SG?

No. The method is implemented entirely in software — there is no need to replace processors or chips, which dramatically lowers the barrier to adoption.