Don't Paint Everyone with the Same Brush: Adaptive Prompt Prototype Learning for Vision-Language Models

Zhi Chen

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.

Keywords: Vision Language Models, Prototype Learning, Prompt Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Vision Language Models (VLMs) have demonstrated great potential on zero-shot classification tasks by computing the similarity between visual and textual embeddings. To adapt VLMs to a downstream task, recent advances introduced context optimization. It optimizes a single embedding for either visual or textual modalities, aiming to improve performance on both base and new classes. However, we identify a critical issue by using single embedding for each class. That is, for image samples of a single class, the visual appearance may vary significantly. Thus, existing methods relying on a singular textual embedding fail to capture the visual variance, leading to suboptimal performance on downstream tasks. In this paper, we propose an Adaptive Prompt Prototype Learning (APPLe) for VLMs. Specifically, we build various prompts as class prototypes to cover the visual variance. Moreover, there are inevitably some ambiguous words in prompts, bringing noise to the textual features. To resolve this problem, an adaptive attention mechanism is designed to weigh the importance of different prototypes. It learns to assign higher scores to the representative prototypes, and lower scores to the flawed or less representative prototypes. To evaluate the effectiveness of APPLe, we conduct experiments on three representative tasks, i.e., generalization to unseen classes, new target datasets, and unseen domain shifts. APPLe exhibits a consistent performance improvement of 3.66% on new classes and 2.79% on the harmonic mean.

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: 1016