Missing-Modality Learning
Overview
Missing-modality learning studies how to make systems use information from multiple sources—such as images and clinical records, audio and video, or different sensors—even when some inputs are unavailable. Common approaches simulate missing inputs with modality dropout, adapt fusion according to availability or reliability, or reconstruct missing features using generative models, prototypes, retrieval, and cross-modal distillation; some methods avoid reconstruction and fuse only observed inputs. Research spans medical imaging and electronic health records, sentiment analysis, remote sensing, and robotics, with evaluations increasingly testing varied missing patterns and corruption levels. A central tension is that reconstruction can add useful information but can also introduce errors, so systems must learn when to repair, ignore, or downweight missing or unreliable evidence while maintaining calibration and generalization.
Latest papers
187 papers12 papers in the last four weeks, down 48% on the four weeks before. 0.1% of all new papers.