Parallel SGD: When does averaging help?
This work addresses the problem of optimizing parallel SGD for researchers and practitioners, but it is incremental as it builds on existing averaging practices.
The paper investigates how model averaging frequency in parallel SGD affects convergence, showing that for convex objectives it depends on gradient variance envelope and for non-convex objectives on multiple globally optimal points, with experiments on synthetic and real data.
Consider a number of workers running SGD independently on the same pool of data and averaging the models every once in a while -- a common but not well understood practice. We study model averaging as a variance-reducing mechanism and describe two ways in which the frequency of averaging affects convergence. For convex objectives, we show the benefit of frequent averaging depends on the gradient variance envelope. For non-convex objectives, we illustrate that this benefit depends on the presence of multiple globally optimal points. We complement our findings with multicore experiments on both synthetic and real data.