Incrementally Adapting Generative Vision-Language Models with Task Codebook

Jinghao Zhou, Ahmet Iscen, Mathilde Caron, Christian Rupprecht, Philip Torr, Cordelia Schmid

Keywords: Generative Vision-Language Models, Incremental Learning

Abstract:

With the help of large-scale pre-training, generative Vision-Language Models (VLMs) have acquired general-purpose capabilities.
As downstream applications diversify, it is imperative for VLMs to learn and adapt continuously without experiencing catastrophic forgetting or necessitating complete retraining.
In this work, we analyze the forgetting behavior of VLMs and propose a solution to enhance their incremental learning abilities.
We introduce a Task Codebook within VLMs, enabling efficient retrieval of task-specific parameters for model adaptation.
Our evaluation encompasses a diverse set of tasks spanning a wide range of visual domains and textual instructions.
Experiments demonstrate that our approach effectively mitigates forgetting, even under highly demanding task sequences.

Primary Area: transfer learning, meta learning, and lifelong 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: 7645