TiMix: Text-aware Image Mixing for Effective Vision-Language Pre-training
This work addresses data efficiency for vision-language pre-training models, offering a computationally viable solution that could benefit broader adoption in practical scenarios, though it is incremental as it builds on existing mix-based augmentation techniques.
The paper tackles the problem of data inefficiency and high computational cost in scaling up vision-language pre-training (VLP) by proposing TiMix, a text-aware image mixing method that integrates mix-based data augmentation into self-supervised multi-modal contrastive learning, achieving comparable performance on downstream tasks with reduced training data and shorter training time.
Self-supervised Multi-modal Contrastive Learning (SMCL) remarkably advances modern Vision-Language Pre-training (VLP) models by aligning visual and linguistic modalities. Due to noises in web-harvested text-image pairs, however, scaling up training data volume in SMCL presents considerable obstacles in terms of computational cost and data inefficiency. To improve data efficiency in VLP, we propose Text-aware Image Mixing (TiMix), which integrates mix-based data augmentation techniques into SMCL, yielding significant performance improvements without significantly increasing computational overhead. We provide a theoretical analysis of TiMixfrom a mutual information (MI) perspective, showing that mixed data samples for cross-modal contrastive learning implicitly serve as a regularizer for the contrastive loss. The experimental results demonstrate that TiMix exhibits a comparable performance on downstream tasks, even with a reduced amount of training data and shorter training time, when benchmarked against existing methods. This work empirically and theoretically demonstrates the potential of data mixing for data-efficient and computationally viable VLP, benefiting broader VLP model adoption in practical scenarios.