Using geometric structure of data for multimodal manifold learning