Efficient Compression of Overparameterized Deep Models through Low-Dimensional Learning Dynamics

The author proposes a novel approach to compress overparameterized deep models by studying their learning dynamics, leading to faster convergence and reduced training costs.