FFF: Fixing Flawed Foundations in contrastive pre-training results in very strong Vision-Language models
This work improves vision-language models for tasks like image recognition and retrieval, but it is incremental as it builds on known bottlenecks in contrastive pre-training.
The paper tackled issues of noise and caption quality in vision-language contrastive pre-training by addressing incorrect negative pair assignments and low caption quality/diversity, resulting in large gains such as ~+6% on average over 11 datasets for image recognition and ~+19% on Flickr30k for image retrieval.
Despite noise and caption quality having been acknowledged as important factors impacting vision-language contrastive pre-training, in this paper, we show that the full potential of improving the training process by addressing such issues is yet to be realized. Specifically, we firstly study and analyze two issues affecting training: incorrect assignment of negative pairs, and low caption quality and diversity. Then, we devise effective solutions for addressing both problems, which essentially require training with multiple true positive pairs. Finally, we propose training with sigmoid loss to address such a requirement. We show very large gains over the current state-of-the-art for both image recognition ($\sim +6\%$ on average over 11 datasets) and image retrieval ($\sim +19\%$ on Flickr30k and $\sim +15\%$ on MSCOCO).