We present a numerical method for learning the dynamics of slow components of unknown Multiscale stochastic dynamical systems. While the governing equations of the systems are unknown, bursts of observation data of the slow variables are available. By utilizing the observation data, ou