Salvato in:
Dettagli Bibliografici
Autori principali: Chen, Liang, Ahmad, Ghazi Shazan, Yao, Tianjun, Liu, Lingqiao, Shen, Zhiqiang
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2507.15480
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916866734161920
author Chen, Liang
Ahmad, Ghazi Shazan
Yao, Tianjun
Liu, Lingqiao
Shen, Zhiqiang
author_facet Chen, Liang
Ahmad, Ghazi Shazan
Yao, Tianjun
Liu, Lingqiao
Shen, Zhiqiang
contents Pretrained vision-language models (VLMs), such as CLIP, achieve remarkable zero-shot performance, yet their downstream potential hinges on effective fine-tuning. Most adaptation methods typically focus on refining representation from separate modalities (text or vision) but neglect the critical role of their fused representations in the decision-making process, \emph{\ie} rational matrix that drives the final prediction. To bridge the gap, we propose a simple yet effective \textbf{R}ational \textbf{Ada}ptaion ({RAda}) to explicitly exploit the final fused representation during fine-tuning. RAda employs a learned mask, obtained from a lightweight attention layer attached at the end of a VLM, to dynamically calibrate the contribution of each element in the rational matrix, enabling targeted adjustments to the final cross-modal interactions without incurring costly modifications to intermediate features. Experiments in different settings (i.e., updating, or freezing pretrained encoders in adaptation, and test-time training that can only access the unlabeled test data) show that RAda serves as a versatile fine-tuning technique, improving the baseline with minimal code and performing comparably against current arts in most settings. Code is available at \href{https://github.com/khufia/RAda/tree/main}{github.com/khufia/RAda}.
format Preprint
id arxiv_https___arxiv_org_abs_2507_15480
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle One Last Attention for Your Vision-Language Model
Chen, Liang
Ahmad, Ghazi Shazan
Yao, Tianjun
Liu, Lingqiao
Shen, Zhiqiang
Computer Vision and Pattern Recognition
Pretrained vision-language models (VLMs), such as CLIP, achieve remarkable zero-shot performance, yet their downstream potential hinges on effective fine-tuning. Most adaptation methods typically focus on refining representation from separate modalities (text or vision) but neglect the critical role of their fused representations in the decision-making process, \emph{\ie} rational matrix that drives the final prediction. To bridge the gap, we propose a simple yet effective \textbf{R}ational \textbf{Ada}ptaion ({RAda}) to explicitly exploit the final fused representation during fine-tuning. RAda employs a learned mask, obtained from a lightweight attention layer attached at the end of a VLM, to dynamically calibrate the contribution of each element in the rational matrix, enabling targeted adjustments to the final cross-modal interactions without incurring costly modifications to intermediate features. Experiments in different settings (i.e., updating, or freezing pretrained encoders in adaptation, and test-time training that can only access the unlabeled test data) show that RAda serves as a versatile fine-tuning technique, improving the baseline with minimal code and performing comparably against current arts in most settings. Code is available at \href{https://github.com/khufia/RAda/tree/main}{github.com/khufia/RAda}.
title One Last Attention for Your Vision-Language Model
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2507.15480