Part of Advances in Neural Information Processing Systems 32 (NeurIPS 2019)
Pooria Joulani, András György, Csaba Szepesvari
We present two new algorithms, ASYNCADA and HEDGEHOG, for asynchronous sparse online and stochastic optimization. ASYNCADA is, to our knowledge, the first asynchronous stochastic optimization algorithm with finite-time data-dependent convergence guarantees for generic convex constraints. In addition, ASYNCADA: (a) allows for proximal (i.e., composite-objective) updates and adaptive step-sizes; (b) enjoys any-time convergence guarantees without requiring an exact global clock; and (c) when the data is sufficiently sparse, its convergence rate for (non-)smooth, (non-)strongly-convex, and even a limited class of non-convex objectives matches the corresponding serial rate, implying a theoretical “linear speed-up”. The second algorithm, HEDGEHOG, is an asynchronous parallel version of the Exponentiated Gradient (EG) algorithm for optimization over the probability simplex (a.k.a. Hedge in online learning), and, to our knowledge, the first asynchronous algorithm enjoying linear speed-ups under sparsity with non-SGD-style updates. Unlike previous work, ASYNCADA and HEDGEHOG and their convergence and speed-up analyses are not limited to individual coordinate-wise (i.e., “box-shaped”) constraints or smooth and strongly-convex objectives. Underlying both results is a generic analysis framework that is of independent interest, and further applicable to distributed and delayed feedback optimization