Multimodal Disentangled Representation Learning

This field studies how to represent information from sources such as images and text, MRI and PET, or audio and video without losing either cross-modal meaning or each source’s distinctive signals. Common approaches use shared and private latent spaces, variational autoencoders, contrastive objectives, orthogonality or mutual-information constraints, and attention or gated fusion. Researchers evaluate these methods on tasks including sentiment analysis, medical prediction and image translation, retrieval, and robustness to missing modalities or new domains. A central tension is that stronger alignment can suppress useful modality-specific information, while the benefits of disentanglement vary across tasks, architectures, and evaluation measures.