The Neglected Tails of Vision-Language Models

Vision-language models (VLMs) excel in zero-shot recognition but exhibit drastically imbalanced performance across visual concepts. For example, CLIP, despite an impressive mean zero-shot accuracy on ImageNet (72.7%), yields $<$10% on ten concepts (e.g., gyromitra and night snake), presumably, because these concepts are under-represented in VLMs' imbalanced pretraining data. Yet, assessing this imbalance is challenging as it is non-trivial to calculate the frequency of specific concepts within VLMs' large-scale pretraining data. Our work makes the first attempt to measure the concept frequency by analyzing pretraining texts. We use off-the-shelf language models to help count relevant texts that contain synonyms of the given concepts and resolve linguistic ambiguity. We confirm that popular VLM datasets like LAION indeed exhibit long-tailed concept distributions, which strongly correlate with per-class accuracies. Further, contemporary multimodal systems, e.g., visual chatbots and text-to-image generators, also struggle with the rare concepts identified by our method. To mitigate VLMs' imbalanced performance in zero-shot recognition, we propose REtrieval-Augmented Learning REAL. First, instead of prompting VLMs using the original class names, REAL uses their most frequent synonyms found in VLMs' pretraining texts. This already outperforms human-engineered and LLM-generated prompts over nine benchmark datasets, likely because VLMs have seen more images associated with the frequently used synonyms. Second, REAL uses all the concept synonyms to retrieve a small, class-balanced set of pretraining data to train a robust classifier. REAL surpasses the recent retrieval-augmented solution REACT, using 400x less storage and 10,000x less training time!