Fast simulations of differential equations(DE) are crucial for ensemble predictions in uncertainty quantification, with many applications ranging from weather forecasting to finance. However, classical numerical methods are often computationally prohibitive for such tasks, because these DEs are often high-dimensional and stiff. Data-driven reduced models have succeeded in such tasks by focusing computational resources on the quantities of interest with the help of data. In this talk, I will cast the data-driven reduced model as an inference-based approximation of the discrete-time flow map for the dynamics of the variables of interest. This flow map approximation framework presents fruitful connections between classical numerical methods and data-driven machine learning methods.