KIST Lifts Neuromorphic AI to ImageNet Lead With 1/6 of Google Training Cost
Updated
Updated · Newswise · Aug 10
KIST Lifts Neuromorphic AI to ImageNet Lead With 1/6 of Google Training Cost
1 articles · Updated · Newswise · Aug 10
Summary
KIST said its new A²SG learning method pushed spiking neural networks to world-leading ImageNet image-recognition accuracy, a milestone for low-power neuromorphic AI.
A²SG improves training by combining adaptive tuning with an asymmetric neuron-inspired approach, addressing the gradient limitations that have kept SNNs behind conventional deep neural networks.
The method delivered higher accuracy with about one-sixth the computational overhead of Google's leading training approach and worked across small and large models, including transformer-based SNNs.
Accepted as a regular paper at ICML 2026 in Seoul, the work gives South Korea a stronger foothold in AI semiconductor technology that had been led largely by universities and global tech firms.
Because A²SG requires only software changes, KIST said it could be deployed in on-device AI for smartphones, wearables, drones and always-on sensors before later commercialization with neuromorphic hardware.