A research team led by Professor Paik Se-bum at KAIST's Cognitive Intelligence Laboratory demonstrated that pre-training neural networks with random noise, similar to biological brains, could match the learning efficiency of conventional AI systems.
The breakthrough offers an alternative to the error backpropagation learning method, first proposed in 1986 by 2024 Nobel Physics laureate Geoffrey Everest Hinton, which requires extensive computational resources and large-scale data processing.
"By solving the weight transport problem in machine learning, we have presented a new perspective that could bridge the gap between artificial neural networks and brain learning," Paik said. "This could contribute to next-generation AI development."
The findings will be presented at the 38th Conference on Neural Information Processing Systems in Vancouver, Canada, Dec. 10-15.
