Excess-Risk Analysis of Distributed Stochastic Learners

In this work, we analyze the learning ability of diffusion-based distributed learners that receive a continuous stream of data arising from the same distribution. We establish four distinctive advantages for these learners relative to other decentralized schemes. First, we obtain closed-form expressions for the evolution of their excess-risk for strongly-convex risk functions under a diminishing step-size rule. Using the result, we then show that the distributed strategy can improve the asymptotic convergence rate of the excess-risk by a factor of N relative to non-cooperative schemes, where N is the number of learners in the ad-hoc network. We further show that the fastest attainable rate of convergence matches the Cramer-Rao bound (up to constants that do not depend on N or the iteration number i) under some mild regularity conditions on the distribution of the data. Finally, we show that the diffusion strategy outperforms consensus-based strategies by reducing the overshoot during the transient phase of the learning process and asymptotically as well. In light of these properties, diffusion strategies are shown to enhance the learning ability of ad-hoc distributed networks by relying solely on localized interactions and on in-network processing.