Multi-modal Domain Adapter Through Hyperbolic and Topic Space
Keywords: Multi-modal search, Hyperbolic space, Hyperbolic geometry, Lorentz model, Domain adapter
TL;DR: Chimera is an adaptation framework that enhances their representation learning over modalities.
Abstract: As multi-modal search relies on jointly learning image-text representations and has been investigated in the literature,
we explore spaces or representations that coexist with and enhance these models.
Because the core of multi-modal search is learning their modalities in a shared semantic space and measuring their similarities,
search quality depends on which expressive space is used for learning.
We find that topic and Hyperbolic space can complement this space,
and propose a model-agnostic adaptation framework, Chimera.
The novelty of this framework lies in 1) designing the topic and Hyperbolic spaces to reveal relationships buried in traditional spaces,
2) leveraging token level interactions,
and 3) performing contrastive/metric learning tasks to ensure the cooperation of these spaces with pre-trained models.
Experiments show that Chimera empowers pre-trained models for multi-modal search tasks and demonstrate the ability of the layers it introduces.
Primary Area: representation learning for computer vision, audio, language, and other modalities
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/2024/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: 9317
Loading