Embedding-Converter: A Unified Framework for Cross-Model Embedding Transformation

Jinsung Yoon, Sercan O Arik

Keywords: Embeddings, Embedding Converter, Embedding transformation

Abstract: Embeddings, numerical representations of data like text and images, are fundamental to machine learning. However, the continuous emergence of new embedding models poses a challenge: migrating to these potentially superior models often requires computationally expensive re-embedding of entire datasets even without guarantees of improvement. This paper introduces Embedding-Converter, a unified framework and a novel paradigm for efficiently converting embeddings between different models, eliminating the need for costly re-embedding. In real-world scenarios, the proposed method yields O(100) times faster and cheaper computation of embeddings with new models. Our experiments demonstrate that Embedding-Converter not only facilitates seamless transitions to new models but can even surpass the source model's performance, approaching that of the target model. This enables efficient evaluation of new embedding models and promotes wider adoption by reducing the overhead associated with model switching. Moreover, Embedding-Converter addresses latency constraints by enabling the use of smaller models for online tasks while leveraging larger models for offline processing. By encouraging users to release converters alongside new embedding models, Embedding-Converter fosters a more dynamic and user-friendly paradigm for embedding model development and deployment.

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.

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: 3829