Factorized contrastive learning is a self-supervised representation learning framework that decomposes multimodal data into distinct shared and modality-unique representations. Unlike standard multimodal contrastive methods that capture only the redundant information overlapping across paired modalities, factorized contrastive learning explicitly isolates features common to all modalities from features unique to individual modalities. By optimizing information-theoretic objectives to preserve task-relevant shared and modality-specific signals while discarding irrelevant noise, this approach enables machine learning models to capture both cross-modal correlations and complementary, single-modality information for downstream tasks.
