Multi-View Learning

Multi-view learning studies how to learn from several representations of the same objects, such as camera angles, sensors, modalities, or feature spaces, to use shared evidence and complementary signals. Common approaches include canonical correlation and joint-individual factor models, contrastive and self-supervised alignment, shared/private representation learning, and graph- or attention-based fusion; optimal transport can help align views with uncertain correspondences. Applications range from 3D perception and robotics to medical imaging, brain data, recommendation, and multi-omics, with evaluation often testing missing views, noise, domain shifts, or limited labels. A central challenge is deciding when and how to fuse: views may be redundant, unreliable, conflicting, or unevenly informative, so more inputs do not always help; current work examines view selection, uncertainty-aware fusion, incomplete data, and theoretical guarantees.