Understanding Retrieval Robustness for Retrieval-Augmented Image Captioning

TL;DR

This paper analyzes the robustness of retrieval-augmented image captioning models, revealing their sensitivity to common tokens in retrieved captions and proposing a training method to improve diversity and performance.

Contribution

It identifies the vulnerability of retrieval-augmented models to majority tokens and introduces a training approach that enhances robustness and cross-domain effectiveness.

Findings

Model is sensitive to tokens appearing in most retrieved captions.

Tokens are often copied into generated outputs.

Diverse sampling during training improves robustness.

Abstract

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.