查看文章

In mobile multimedia applications, deep learning has received significant interest. Due to the limited computation and storage resources of mobile devices, however, general training methods are hardly suited for mobile multimedia computing. For this reason, we propose an adaptive momentum training (FWAdaBound) algorithm to reduce computation and storage cost, where the Frank‐Wolfe method is employed. Furthermore, we rigorously prove the regret bound in order that O(T^3/4) can be achieved, where T is a time horizon. Finally, the convergence, cost‐reduction, and generalization ability of FWAdaBound are validated through various experiments on public datasets.