Weather and climate models approximate diabatic and sub-grid-scale processes in terms of grid-scale variables using parameterizations. Current parameterizations are de- signed by humans based on physical understanding, observations and process modeling. As a result, they are numerically efficient and interpretable, but potentially over-simplified. However, the advent of global high-resolution simulations and observations enables a more robust approach based on machine learning. In this letter, a neural network (NN) based parameterization is trained using a global-scale cloud-resolving simulation. The NN predicts the apparent sources of heat and moisture averaged onto (160 km)^2 grid boxes. A numerically stable scheme is obtained by minimizing the prediction error over multiple time steps rather than single one. In prognostic single column model tests, this scheme outperforms the Community Atmosphere Model by reducing both long-term bias and short-term errors.