Dataset distillation methods promise to reduce large-scale datasets down to significantly smaller sets of (potentially synthetic) training examples, which preserve sufficient information for training a new model from scratch. So far, dataset distillation methods have been developed for image classification. However, with the rise in capabilities of vision-language models (VLMs), and especially given the scale of datasets necessary to train these models, the time is ripe to expand dataset distillation methods beyond image classification. In this work, we take the first steps towards this goal by expanding the idea of trajectory matching to create a distillation method for vision-language datasets. A key challenge is that vision-language datasets do not have a set of discrete classes. To overcome this, our proposed vision-language dataset distillation method jointly distills the image-text pairs in a contrastive formulation. Since there are no existing baselines, we compare our approach to three coreset selection methods (strategic subsampling of the training dataset), which we adapt to the vision-language setting. We demonstrate significant improvements on the challenging Flickr30K and COCO retrieval benchmarks: for example, on Flickr30K, the best coreset selection method selecting 1000 image-text pairs for training achieves only 5.6% image-to-text retrieval accuracy (i.e., recall@1); in contrast, our dataset distillation approach almost doubles that to 9.9% with just 100 (an order of magnitude fewer) training pairs.
View on arXiv