We have applied this idea to compute Margin Valuation Adjustments (MVA) for an interest rate swap book, in a use case for a major bank. Recall that computing margin valuation adjustment involves predicting future Initial Margin (IM) charged by a clearing house on (a portfolio of) derivatives, and these, in turn, are computed using the sensitivities of these derivatives to underlying factors (such as interest rates or volatilities). It is the precise calculation of these sensitivities that is usually cumbersome and time consuming. For example, numerical differentiation algorithms (the so-called "bump-and-revalue") are inefficient and inaccurate. The modern technique of Algorithmic Adjoint Differentiation (AAD) does significantly better on both fronts, but can be still quite time-consuming (and complex). So we generated a large number of interest rate curves scenarios (100 000) in the two-curves setting – we used the two-factor Hull-White model – and then trained a deep learning algorithm (Feed-Forward Neural Network) to approximate the sensitivities of interest rate derivatives to the swap rates as a function of maturity, volatility and other parameters.