UnICLAM: Contrastive Representation Learning with Adversarial Masking for Unified and Interpretable Medical Vision Question Answering.