Provable Benefit of Orthogonal Initialization in Optimizing Deep Linear Networks

Open in new window