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