Towards Multimodal Open-World Learning In Deep Neural Networks