Multimodal Generative Learning Utilizing Jensen-Shannon-Divergence