Leveraging hierarchy in multimodal generative models for effective cross-modality inference

This work addresses the problem of cross-modality inference (CMI), i.e., inferring missing data of unavailable perceptual modalities (e.g., sound) using data from available perceptual modalities (e.g., image). We overview single-modality variational autoencoder methods and discuss three problems of computational cross-modality inference, arising from recent developments in multimodal generative models. Inspired by neural mechanisms of human recognition, we contribute the Nexus model, a novel hierarchical generative model that can learn a multimodal representation of an arbitrary number of modalities in an unsupervised way. By exploiting hierarchical representation levels, Nexus is able to generate high-quality, coherent data of missing modalities given any subset of available modalities. To evaluate CMI in a natural scenario with a high number of modalities, we contribute the "Multimodal Handwritten Digit" (MHD) dataset, a novel benchmark dataset that combines image, motion, sound and label information from digit handwriting. We access the key role of hierarchy in enabling high-quality samples during cross-modality inference and discuss how a novel training scheme enables Nexus to learn a multimodal representation robust to missing modalities at test time. Our results show that Nexus outperforms current state-of-the-art multimodal generative models in regards to their cross-modality inference capabilities.