Zhao Song

The University of Texas at Austin, Simon Fraser University

Papers

Total Citations

178

H-Index

About

Zhao Song is a leading theorist in deep learning, best known for pioneering the convergence theory of over-parameterized neural networks. His seminal 2018 paper, "A Convergence Theory for Deep Learning via Over-Parameterization," with 176 citations, established a rigorous mathematical framework proving that sufficiently wide networks can converge to global minima under gradient descent—a foundational result that demystified why deep networks train successfully in practice. This work has become a cornerstone of modern deep learning theory, influencing subsequent research on optimization, generalization, and neural tangent kernels. Earlier in his career, Song explored multi-agent robotics with "MO-LOST: adaptive ant trail untangling in multi-objective multi-colony robot foraging" (2012), demonstrating versatility in applying biological principles to swarm intelligence. His contributions bridge the gap between theoretical guarantees and empirical deep learning success, offering students and researchers a clear mathematical lens for understanding why over-parameterization works. Song’s research continues to shape how we analyze and design neural networks, making him a key figure in the theoretical foundations of modern AI.