Self-Training on Unpaired Data Improves Multi-Modal Alignment
Keywords: Self-Training, Multi-Modal Representation Alignment
TL;DR: Alternatively self-training on paired data and unpaired data can improves multi-modal alignment
Abstract: In the past few years, multimodal foundation models, e.g., CLIP, learned from a massive amount of paired multimodal data, emerged and exhibited impressive cross-modal ability in many applications. Yet collecting high-quality paired data is generally costly or even infeasible in certain cases, and the amount of paired multimodal data is several orders fewer than that of unpaired unimodal data, i.e., data without any correspondence. Our work focuses on alleviating the excessive demand for paired language-image data by leveraging the abundant unpaired data. We introduce a new approach for vision-language alignment, which we call Language-Image Self-Training (LIST). LIST consists of two key ingredients that function in a synergistic loop: i) a captioner model trained alternatively with the augmented paired data and the unpaired data with synthetic captions, both derived from the data engine, and ii) a data engine that synthesizes a diverse spectrum of captions for both paired and unpaired images with the captioner, integrating synthetic captions with the web-scraped ones to enhance the quality of paired data using off-the-shelf Large Language Models. We observe that the LIST methodology not only significantly improves the alignment between vision and language representations across multiple major benchmarks—zero-shot image classification, image-text retrieval, and compositional evaluation—but also demonstrates strong generalization to audio-language representation alignment.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3843
Loading