Understanding Retrieval Robustness for Retrieval-Augmented Image Captioning
This addresses robustness issues in retrieval-augmented image captioning for AI applications, but it is incremental as it builds on existing models like SmallCap.
The paper tackles the problem of retrieval models misleading retrieval-augmented image captioning models, causing incorrect generation and worse performance, and finds that training with more diverse retrieved captions reduces copying of majority tokens and improves in-domain and cross-domain performance.
Recent advances in retrieval-augmented models for image captioning highlight the benefit of retrieving related captions for efficient, lightweight models with strong domain-transfer capabilities. While these models demonstrate the success of retrieval augmentation, retrieval models are still far from perfect in practice: the retrieved information can sometimes mislead the model, resulting in incorrect generation and worse performance. In this paper, we analyze the robustness of a retrieval-augmented captioning model SmallCap. Our analysis shows that the model is sensitive to tokens that appear in the majority of the retrieved captions, and the input attribution shows that those tokens are likely copied into the generated output. Given these findings, we propose to train the model by sampling retrieved captions from more diverse sets. This decreases the chance that the model learns to copy majority tokens, and improves both in-domain and cross-domain performance.