Distilling Cross-Modal Knowledge via Feature Disentanglement
This addresses the challenge of compressing models in cross-modal applications like vision-to-language, though it is incremental as it builds on existing knowledge distillation techniques.
The paper tackles the problem of knowledge distillation in cross-modal scenarios, where representation inconsistencies hinder effective transfer, by proposing a frequency-decoupled method that applies distinct losses to low- and high-frequency features, resulting in substantial performance improvements over traditional and state-of-the-art approaches across multiple benchmark datasets.
Knowledge distillation (KD) has proven highly effective for compressing large models and enhancing the performance of smaller ones. However, its effectiveness diminishes in cross-modal scenarios, such as vision-to-language distillation, where inconsistencies in representation across modalities lead to difficult knowledge transfer. To address this challenge, we propose frequency-decoupled cross-modal knowledge distillation, a method designed to decouple and balance knowledge transfer across modalities by leveraging frequency-domain features. We observed that low-frequency features exhibit high consistency across different modalities, whereas high-frequency features demonstrate extremely low cross-modal similarity. Accordingly, we apply distinct losses to these features: enforcing strong alignment in the low-frequency domain and introducing relaxed alignment for high-frequency features. We also propose a scale consistency loss to address distributional shifts between modalities, and employ a shared classifier to unify feature spaces. Extensive experiments across multiple benchmark datasets show our method substantially outperforms traditional KD and state-of-the-art cross-modal KD approaches. Code is available at https://github.com/Johumliu/FD-CMKD.