Improved Multimodal Deep Learning with Variation of Information