Rest-Katyusha: Exploiting the Solution's Structure via Scheduled Restart Schemes
Tang, Junqi, Golbabaee, Mohammad, Bach, Francis, davies, Mike E.
–Neural Information Processing Systems
We propose a structure-adaptive variant of the state-of-the-art stochastic variance-reduced gradient algorithm Katyusha for regularized empirical risk minimization. The proposed method is able to exploit the intrinsic low-dimensional structure of the solution, such as sparsity or low rank which is enforced by a non-smooth regularization, to achieve even faster convergence rate. This provable algorithmic improvement is done by restarting the Katyusha algorithm according to restricted strong-convexity constants. We demonstrate the effectiveness of our approach via numerical experiments.
Neural Information Processing Systems
Dec-31-2018