Rest-Katyusha: Exploiting the Solution's Structure via Scheduled Restart Schemes

We propose a structure-adaptive variant of the state-of-the-art stochastic variance-reduced gradient algorithm Katyusha for the regularized empirical risk minimization. The proposed method is able to exploit the intrinsic low-dimensional structure of the solution, such as sparsity and low-rank, which is enforced by the non-smooth regularization, to achieve even faster convergence rate. This algorithmic improvement is done by restarting the Katyusha algorithm at a certain carefully-chosen frequency according to a modified version of restricted strong-convexity. Our analysis demonstrates that the proposed method is globally convergent and enjoys a local accelerated linear rate with respect to the low-dimensional structure of the solution represented by the restricted strong-convexity, even when the cost function itself is not strongly-convex. Since in practice the restricted strong-convexity is usually unknown and hard to be estimated accurately, we proposed two practical restart schemes. The first one is restarting the algorithm with a rough restricted strong-convexity estimate which is provably robust but have a compromise on the convergence rate. The second variant is based on the adaptive restart via convergence speed check. The numerical results on benchmark datasets demonstrate the effectiveness of our approach.

Paper

Similar papers

© 2026 NYSGPT2525 LLC