Brain-inspired sparse network science for efficient and sustainable AI
COM3 Level 2
MR21, COM3 02-61
Abstract
Training large language models can require gigawatts of power, whereas the human brain operates at roughly 20 watts. Recent theoretical work suggests that dense neural networks are not universal approximators, pointing instead to sparse connectivity as a necessary ingredient for universality. Brain-inspired network science offers a principled path toward low-consumption, efficient deep learning, supporting a more sustainable approach to AI by drawing on the physics of brain network architecture and the principles of complex systems biology. At the Center for Complex Network Intelligence (CCNI) within the Tsinghua Laboratory of Brain and Intelligence (THBI), research focuses on three features of brain networks that enable efficient computation: connectivity sparsity, which reduces computational cost while preserving performance; connectivity morphology, which studies spatial organization to improve information processing; and neuro-glia coupling, which examines neuron–glia interactions for enhanced efficiency. This talk introduces the Cannistraci–Hebb Training soft rule (CHTs), a brain-inspired network science framework that uses a gradient-free approach, relying on network topology to guide sparse connectivity during dynamic training, and demonstrates an ultra-sparse regime—around 1% connectivity—that can outperform fully connected networks across a range of tasks.
Bio
Carlo Vittorio Cannistraci is a theoretical engineer and computational innovator. He is Chair Professor at the Tsinghua Laboratory of Brain and Intelligence (THBI) and holds appointments in the Department of Psychological and Cognitive Sciences, the Department of Computer Science and Technology, and the School of Biomedical Engineering at Tsinghua University. He directs the Center for Complex Network Intelligence (CCNI) at THBI, where research develops algorithms at the interface of information science, complex systems physics, network science, and machine intelligence, with an emphasis on brain- and life-inspired computing for efficient AI and large-scale data analysis. These methods are applied to precision biomedicine, neuroscience, and the social and economic sciences.