Multimodal representation and learning

Recent years have seen an explosion in multimodal data on the web. It is therefore important to perform multimodal learning to understand the web. However, it is challenging to join various modalities because each modality has a different representation and correlational structure. In addition, various modalities generally carry different kinds of information that may provide enrich understanding; for example, the visual signal of a flower may provide happiness; however, its scent might not be pleasant. Multimodal information may be useful to make an informed decision. Therefore, we focus on improving representations from individual modalities to enhance multimodal representation and learning. In this doctoral thesis, we presented techniques to enhance representations from individual and multiple modalities for multimodal applications including classification, cross-modal retrieval, matching and verification on various benchmark datasets.