Reproduce “FILIP: Fine-grained Interactive Language-Image Pre-Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li and 2 more
FILIP, developed by Huawei Noah’s Ark Lab and academic partners, introduces a dual-stream vision-language pre-training model that efficiently integrates fine-grained cross-modal interaction. It achieves an average 78.3% top-1 accuracy on 12 zero-shot image classification datasets, outperforming CLIP by 3.0%, and surpasses ALIGN on image-text retrieval tasks while demonstrating implicit word-patch alignment.
Key insights
- Fine-grained cross-modal interaction can be effectively integrated into an efficient dual-stream VLP framework by modifying the contrastive loss, bridging the gap between efficiency and granular understanding.
- Aggregating token-wise maximum similarities by averaging, rather than summing, provides more stable training and better performance by normalizing for varying sequence lengths.
- Computational optimizations, such as reduced embedding dimensions and token subsampling, are crucial for making token-level interactions practical and scalable in large-scale pre-training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.