BayesAdapter: enhanced uncertainty estimation in CLIP few-shot adaptation

-д хадгалсан:
Номзүйн дэлгэрэнгүй
-д хэвлэсэн:arXiv.org (Dec 12, 2024), p. n/a
Үндсэн зохиолч: Morales-Álvarez, Pablo
Бусад зохиолчид: Christodoulidis, Stergios, Vakalopoulou, Maria, Piantanida, Pablo, Dolz, Jose
Хэвлэсэн:
Cornell University Library, arXiv.org
Нөхцлүүд:
Онлайн хандалт:Citation/Abstract
Full text outside of ProQuest
Шошгууд: Шошго нэмэх
Шошго байхгүй, Энэхүү баримтыг шошголох эхний хүн болох!

MARC

LEADER 00000nab a2200000uu 4500
001 3145272708
003 UK-CbPIL
022 |a 2331-8422 
035 |a 3145272708 
045 0 |b d20241212 
100 1 |a Morales-Álvarez, Pablo 
245 1 |a BayesAdapter: enhanced uncertainty estimation in CLIP few-shot adaptation 
260 |b Cornell University Library, arXiv.org  |c Dec 12, 2024 
513 |a Working Paper 
520 3 |a The emergence of large pre-trained vision-language models (VLMs) represents a paradigm shift in machine learning, with unprecedented results in a broad span of visual recognition tasks. CLIP, one of the most popular VLMs, has exhibited remarkable zero-shot and transfer learning capabilities in classification. To transfer CLIP to downstream tasks, adapters constitute a parameter-efficient approach that avoids backpropagation through the large model (unlike related prompt learning methods). However, CLIP adapters have been developed to target discriminative performance, and the quality of their uncertainty estimates has been overlooked. In this work we show that the discriminative performance of state-of-the-art CLIP adapters does not always correlate with their uncertainty estimation capabilities, which are essential for a safe deployment in real-world scenarios. We also demonstrate that one of such adapters is obtained through MAP inference from a more general probabilistic framework. Based on this observation we introduce BayesAdapter, which leverages Bayesian inference to estimate a full probability distribution instead of a single point, better capturing the variability inherent in the parameter space. In a comprehensive empirical evaluation we show that our approach obtains high quality uncertainty estimates in the predictions, standing out in calibration and selective classification. Our code is publicly available at: https://github.com/pablomorales92/BayesAdapter. 
653 |a Estimates 
653 |a Visual tasks 
653 |a Visual discrimination 
653 |a Classification 
653 |a Parameter estimation 
653 |a Bayesian analysis 
653 |a Probabilistic inference 
653 |a Machine learning 
653 |a Parameter uncertainty 
653 |a Statistical analysis 
653 |a Statistical inference 
653 |a Adapters 
653 |a Back propagation 
700 1 |a Christodoulidis, Stergios 
700 1 |a Vakalopoulou, Maria 
700 1 |a Piantanida, Pablo 
700 1 |a Dolz, Jose 
773 0 |t arXiv.org  |g (Dec 12, 2024), p. n/a 
786 0 |d ProQuest  |t Engineering Database 
856 4 1 |3 Citation/Abstract  |u https://www.proquest.com/docview/3145272708/abstract/embedded/ZKJTFFSVAI7CB62C?source=fedsrch 
856 4 0 |3 Full text outside of ProQuest  |u http://arxiv.org/abs/2412.09718