1: \begin{abstract}

2:  {We propose stochastic modified equations (SMEs) for modeling the asynchronous stochastic gradient

3:   descent (ASGD) algorithms. The resulting SME of Langevin type extracts more information about the

4:   ASGD dynamics and elucidates the relationship between different types of stochastic gradient

5:   algorithms. We show the convergence of ASGD to the SME in the continuous time limit, as well as

6:   the SME's precise prediction to the trajectories of ASGD with various forcing terms. As an

7:   application, we propose an optimal mini-batching strategy for ASGD via solving the

8:   optimal control problem of the associated SME.}

9: %%%% If classification number provided then

10: 

11: \end{abstract}

12: