The artificial intelligence boom has an energy problem, as data centers with power-hungry processors drive demand for generative AI. Researchers are exploring neuromorphic computing, which imitates the human brain, to deliver advanced intelligence with lower energy costs. A research team led by Senior Researcher Seongsik Park of the Semiconductor Technology Research Division at the Korea Institute of Science and Technology (KIST) developed a learning technique called A²SG, or Adaptive and Asymmetric Surrogate Gradients, to improve the training of spiking neural networks, AI models inspired by the brain’s sparse, event-driven communication. Unlike conventional networks that exchange numerical values continuously, spiking neural networks communicate through brief events called spikes, remaining inactive until signals reach a threshold, which can reduce energy consumption in applications like cameras, wearable sensors, drones, and edge devices. However, spiking neurons produce abrupt, discontinuous spikes, making exact gradients zero or undefined and hindering conventional backpropagation. Researchers use surrogate gradients to approximate missing gradients around a neuron’s firing threshold. The KIST team’s A²SG method advances this strategy in two directions. The research was accepted as a regular paper at the International Conference on Machine Learning, or ICML 2026, and presented on July 7 at the conference in Seoul, the first time the event has been held in South Korea.