The robustness of a parallel-jaw grasp can be estimated by Monte Carlo sampling of perturbations in pose and friction but this is not computationally efficient. As an alternative, we consider fast methods using large-scale supervised learning, where the input is a description of a local surface patch at each of two contact points. We train and test with disjoint subsets of a corpus of 1.66 million grasps where robustness is estimated by Monte Carlo sampling using Dex-Net 1.0. We use the BIDMach machine learning toolkit to compare the performance of two supervised learning methods: Random Forests and Deep Learning. We find that both of these methods learn to estimate grasp robustness fairly reliably in terms of Mean Absolute Error (MAE) and ROC Area Under Curve (AUC) on a held-out test set. Speedups over Monte Carlo sampling are approximately 7500x for Random Forests and 1500x for Deep Learning.