Generalization of machine learning models for medical image classification using causal theory. Among other things, I want to investigate the guarantees we can get when there is a distributional shift between training and test data. Further, can we apply techniques (such as privileged information) that lead to better generalization? The resources will primarily be used to train machine learning models on image classification tasks (most likely using popular architectures such as ResNet) and evaluate how these perform under different distributional shifts.