AI RESEARCH
Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification
arXiv CS.CV
•
ArXi:2601.22853v3 Announce Type: replace Multimodal deep learning (MDL) has achieved remarkable success across various domains, yet its practical deployment is often hindered by incomplete multimodal data. Existing incomplete MDL methods either discard missing modalities, risking the loss of valuable task-relevant information, or recover them, potentially