We propose a new framework for regularization, called marginalized corrupted features, that reduces over fitting by increasing the robustness of the model to data corruptions.
