Stability and Generalization of Decentralized Stochastic Gradient Descent

The stability and generalization of stochastic gradient-based methods provide valuable insights into understanding the algorithmic performance machine learning models. As main workhorse for deep learning, gradient descent has received a considerable amount studies. Nevertheless, community paid little attention to its decentralized variants. In this paper, we novel formulation descent. Leveraging together with (non)convex optimization theory, establish first guarantees Our theoretical results are built on top few common mild assumptions reveal that decentralization deteriorates SGD time. We verify our findings by using variety settings benchmark