Multimodal Federated Learning
Overview
Multimodal federated learning studies how institutions or devices can jointly learn from combinations such as images and text, or audio and sensor streams, without pooling their raw data. Common approaches combine local multimodal fusion with federated aggregation, and use representation or prototype alignment, personalized adapters, distillation, and imputation to handle clients with different data distributions or missing modalities. Work spans medical imaging, sentiment analysis, graph learning, activity recognition, and vision-language models, with methods also targeting communication costs and privacy. A central challenge is balancing useful knowledge sharing with client-specific needs when modalities, data quality, and model architectures differ across participants.
Latest papers
51 papers3 papers in the last four weeks, down 40% on the four weeks before. 0.0% of all new papers.