Distilling vision-language pre-training models with modality-specific meta-learning

Abstract

Vision-language pre-training (VLP) models have exhibited excellent performance on diverse vision-language tasks, while the ensuing large-scale model parameters greatly limit their application. Knowledge distillation (KD) makes it possible to apply a VLP model to scenarios with limited resources and real-time responses by transferring the knowledge from it (i.e., teacher) into a lightweight one (i.e., student). However, existing KD methods are primarily designed for unimodal models and thus fail to realize their full potential when migrating them to distill VLP models considering the presence of multiple modalities. Moreover, these KD strategies only unilaterally force the student model to approach the output feature maps generated by the teacher model while ignoring the deeper correlations between them, which may hinder effective knowledge transfer. To tackle these issues, we propose MMKD, a multimodal Knowledge Distillation method with Modality-specific Meta-learning, in which the training objective of the teacher model is converted to optimize the teaching ability for knowledge transfer through feedback from the student model. Meanwhile, to disentangle mutual interference between different modalities when applying KD, the modality-specific distillation objective is designed to encourage the student model to learn the teacher’s knowledge from different modalities. By progressively optimizing the teacher model towards the direction of maximizing the student’s performance, more appropriate soft labels are generated to help the student model learn across different modalities, leading to improved performance. Experiments on three types of VLP models across different downstream tasks demonstrate that the superiority of the proposed method in compressing VLP models.