How Does Fine-Tuning Impact Out-of-Distribution Detection for Vision-Language Models?

被引:4
|
作者
Ming, Yifei [1 ]
Li, Yixuan [1 ]
机构
[1] Univ Wisconsin Madison, Dept Comp Sci, Madison, WI 53715 USA
基金
美国国家科学基金会;
关键词
CLIP; OOD detection; Fine-tuning; Multi-modality; Vision-language models; Prompt learning; Few-shot learning; Adaptor; BLIND DECONVOLUTION; IDENTIFIABILITY; KERNEL; NOISE;
D O I
10.1007/s11263-023-01895-7
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Recent large vision-language models such as CLIP have shown remarkable out-of-distribution (OOD) detection and generalization performance. However, their zero-shot in-distribution (ID) accuracy is often limited for downstream datasets. Recent CLIP-based fine-tuning methods such as prompt learning have demonstrated significant improvements in ID classification and OOD generalization where OOD labels are available. Nonetheless, it remains unclear whether the model is reliable to semantic shifts without OOD labels. In this paper, we aim to bridge the gap and present a comprehensive study to understand how fine-tuning impact OOD detection for few-shot downstream tasks. By framing OOD detection as multi-modal concept matching, we establish a connection between fine-tuning methods and various OOD scores. Our results suggest that a proper choice of OOD scores is essential for CLIP-based fine-tuning. In particular, the maximum concept matching (MCM) score provides a promising solution consistently. We also show that prompt learning demonstrates the state-of-the-art OOD detection performance over the zero-shot counterpart.
引用
收藏
页码:596 / 609
页数:14
相关论文
共 50 条
  • [31] Latent Transformer Models for out-of-distribution detection
    Graham, Mark S.
    Tudosiu, Petru-Daniel
    Wright, Paul
    Pinaya, Walter Hugo Lopez
    Teikari, Petteri
    Patel, Ashay
    U-King-Im, Jean-Marie
    Mah, Yee H.
    Teo, James T.
    Jager, Hans Rolf
    Werring, David
    Rees, Geraint
    Nachev, Parashkev
    Ourselin, Sebastien
    Cardoso, M. Jorge
    [J]. MEDICAL IMAGE ANALYSIS, 2023, 90
  • [32] Deep Hybrid Models for Out-of-Distribution Detection
    Cao, Senqi
    Zhang, Zhongfei
    [J]. 2022 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2022), 2022, : 4723 - 4733
  • [33] Show Me How It's Done: The Role of Explanations in Fine-Tuning Language Models
    Ballout, Mohamad
    Krumnack, Ulf
    Heidemann, Gunther
    Kuehnberger, Kai-Uwe
    [J]. ASIAN CONFERENCE ON MACHINE LEARNING, VOL 222, 2023, 222
  • [34] On the Impact of Spurious Correlation for Out-of-Distribution Detection
    Ming, Yifei
    Yin, Hang
    Li, Yixuan
    [J]. THIRTY-SIXTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE / THIRTY-FOURTH CONFERENCE ON INNOVATIVE APPLICATIONS OF ARTIFICIAL INTELLIGENCE / TWELVETH SYMPOSIUM ON EDUCATIONAL ADVANCES IN ARTIFICIAL INTELLIGENCE, 2022, : 10051 - 10059
  • [35] Knowledge-Aware Prompt Tuning for Generalizable Vision-Language Models
    Kan, Baoshuo
    Wang, Teng
    Lu, Wenpeng
    Zhen, Xiantong
    Guan, Weili
    Zheng, Feng
    [J]. 2023 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV 2023), 2023, : 15624 - 15634
  • [36] Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?
    Wu, Cheng-En
    Tian, Yu
    Yu, Haichao
    Wang, Heng
    Morgado, Pedro
    Hu, Yu Hen
    Yang, Linjie
    [J]. 2023 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV 2023), 2023, : 15442 - 15451
  • [37] Prompt Engineering or Fine-Tuning? A Case Study on Phishing Detection with Large Language Models
    Trad, Fouad
    Chehab, Ali
    [J]. MACHINE LEARNING AND KNOWLEDGE EXTRACTION, 2024, 6 (01): : 367 - 384
  • [38] Getting it right: the limits of fine-tuning large language models
    Browning, Jacob
    [J]. ETHICS AND INFORMATION TECHNOLOGY, 2024, 26 (02)
  • [39] Span Fine-tuning for Pre-trained Language Models
    Bao, Rongzhou
    Zhang, Zhuosheng
    Zhao, Hai
    [J]. FINDINGS OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, EMNLP 2021, 2021, : 1970 - 1979
  • [40] Scaling Federated Learning for Fine-Tuning of Large Language Models
    Hilmkil, Agrin
    Callh, Sebastian
    Barbieri, Matteo
    Sutfeld, Leon Rene
    Zec, Edvin Listo
    Mogren, Olof
    [J]. NATURAL LANGUAGE PROCESSING AND INFORMATION SYSTEMS (NLDB 2021), 2021, 12801 : 15 - 23