Solving prediction games with parallel batch gradient descent

Abstract

Learning problems in which an adversary can perturb instances at application time can be modeled as games with datadependent cost functions. In an equilibrium point, the learner’s model parameters are the optimal reaction to the data generator’s perturbation, and vice versa. Finding an equilibrium point requires the solution of a difficult optimization problem for which both, the learner’s model parameters and the possible perturbations are free parameters. We study a perturbation model and derive optimization procedures that use a single iteration of batch-parallel gradient descent and a subsequent aggregation step, thus allowing for parallelization with minimal synchronization overhead.