Removing Bias in Multi-modal Classifiers: Regularization by Maximizing Functional Entropies

Many recent datasets contain a variety of different data modalities, for instance, image, question, and answer data in visual question answering (VQA).