Cross-Modal Knowledge Distillation without Paired Data: Theoretical Foundation and Algorithm

The article presents a new framework for Cross-Modal Knowledge Distillation (CMKD) that operates without the need for paired data, addressing a significant limitation in existing methods. The proposed approach establishes a cross-modal distributional relationship between teacher and student models, focusing on feature and label alignment to mitigate semantic discrepancies. Experimental results demonstrate substantial performance improvements across various multimodal benchmarks, making this framework valuable for practitioners aiming to leverage knowledge distillation in scenarios where paired datasets are impractical.