View in EDS

Vision-Language Efficient Tuning for Mitigating Catastrophic Forgetting in Multi-Modal Learning.

Saved in:
Bibliographic Details
Title: Vision-Language Efficient Tuning for Mitigating Catastrophic Forgetting in Multi-Modal Learning.
Authors: Wang, Yaoming1 (AUTHOR) wang_yaoming@sjtu.edu.cn, Liu, Yuchen1 (AUTHOR) liuyuchen6666@sjtu.edu.cn, Dai, Wenrui1 (AUTHOR) daiwenrui@sjtu.edu.cn, Zhang, Xiaopeng2 (AUTHOR) zhangxiaopeng12@huawei.com, Li, Han1 (AUTHOR) qingshi9974@sjtu.edu.cn, Li, Chenglin1 (AUTHOR) lcl1985@sjtu.edu.cn, Zou, Junni3 (AUTHOR) zoujunni@sjtu.edu.cn, Tian, Qi2 (AUTHOR) tian.qi1@huawei.com, Xiong, Hongkai1 (AUTHOR) xionghongkai@sjtu.edu.cn
Source: International Journal of Computer Vision. Apr2026, Vol. 134 Issue 4, p1-28. 28p.
Subjects: Orthogonalization, Generalization, Machine learning
Abstract: Parameter-Efficient Tuning (PET) has been widely studied to adapt pretrained vision-language models (VLMs) to various downstream tasks. Most existing PET methods are limited to uni-modal tuning of textual modality and cannot fully exploit the generalization ability of VLMs. Recent multi-modal tuning methods cannot address the catastrophic performance loss in generalizing to unseen classes without incurring significant memory and computation overheads. In this paper, we reveal the multi-modal forgetting issue that violates model coupling for constraining visual features and textual embeddings in simultaneously tuning both modalities, and propose a novel Vision Language Efficient Tuning (VioLET) framework to address the issue. We first propose VioLET-CMG that leverages Collaborative Multi-modal Gradients (CMG) by incorporating an additional visual encoder to provide extra textual gradients to regulate the multi-modal tuning. VioLET-CMG avoids conflicts between visual and textual gradients and guarantee modal coupling with textual gradient orthogonalization. Furthermore, we present VioLET-CarE that achieves quasi-orthogonal textual tuning by Controlling language hyperspherical Energy (CarE) to remove the additional visual encoder for enabling collaborative multi-modal gradients in VioLET-CMG. VioLET-CarE enhances computational efficiency via additive low-rank feature transformation for textual tuning. Experimental results demonstrate that the proposed VioLET framework consistently achieves state-of-the-art performance in new class generalization and few-shot recognition using both ResNet-50 and ViT/B-16 backbones. Remarkably, compared with VioLET-CMG, VioLET-CarE further reduces GPU memory cost by 27% and training times by 36% on average with enhanced generalization ability. [ABSTRACT FROM AUTHOR]
Copyright of International Journal of Computer Vision is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
FullText Text:
  Availability: 0
Header DbId: egs
DbLabel: Engineering Source
An: 192202262
AccessLevel: 6
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 0
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Vision-Language Efficient Tuning for Mitigating Catastrophic Forgetting in Multi-Modal Learning.
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AR" term="%22Wang%2C+Yaoming%22">Wang, Yaoming</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> wang_yaoming@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Liu%2C+Yuchen%22">Liu, Yuchen</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> liuyuchen6666@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Dai%2C+Wenrui%22">Dai, Wenrui</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> daiwenrui@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Zhang%2C+Xiaopeng%22">Zhang, Xiaopeng</searchLink><relatesTo>2</relatesTo> (AUTHOR)<i> zhangxiaopeng12@huawei.com</i><br /><searchLink fieldCode="AR" term="%22Li%2C+Han%22">Li, Han</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> qingshi9974@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Li%2C+Chenglin%22">Li, Chenglin</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> lcl1985@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Zou%2C+Junni%22">Zou, Junni</searchLink><relatesTo>3</relatesTo> (AUTHOR)<i> zoujunni@sjtu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Tian%2C+Qi%22">Tian, Qi</searchLink><relatesTo>2</relatesTo> (AUTHOR)<i> tian.qi1@huawei.com</i><br /><searchLink fieldCode="AR" term="%22Xiong%2C+Hongkai%22">Xiong, Hongkai</searchLink><relatesTo>1</relatesTo> (AUTHOR)<i> xionghongkai@sjtu.edu.cn</i>
– Name: TitleSource
  Label: Source
  Group: Src
  Data: <searchLink fieldCode="JN" term="%22International+Journal+of+Computer+Vision%22">International Journal of Computer Vision</searchLink>. Apr2026, Vol. 134 Issue 4, p1-28. 28p.
– Name: Subject
  Label: Subjects
  Group: Su
  Data: <searchLink fieldCode="DE" term="%22Orthogonalization%22">Orthogonalization</searchLink><br /><searchLink fieldCode="DE" term="%22Generalization%22">Generalization</searchLink><br /><searchLink fieldCode="DE" term="%22Machine+learning%22">Machine learning</searchLink>
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: Parameter-Efficient Tuning (PET) has been widely studied to adapt pretrained vision-language models (VLMs) to various downstream tasks. Most existing PET methods are limited to uni-modal tuning of textual modality and cannot fully exploit the generalization ability of VLMs. Recent multi-modal tuning methods cannot address the catastrophic performance loss in generalizing to unseen classes without incurring significant memory and computation overheads. In this paper, we reveal the multi-modal forgetting issue that violates model coupling for constraining visual features and textual embeddings in simultaneously tuning both modalities, and propose a novel Vision Language Efficient Tuning (VioLET) framework to address the issue. We first propose VioLET-CMG that leverages Collaborative Multi-modal Gradients (CMG) by incorporating an additional visual encoder to provide extra textual gradients to regulate the multi-modal tuning. VioLET-CMG avoids conflicts between visual and textual gradients and guarantee modal coupling with textual gradient orthogonalization. Furthermore, we present VioLET-CarE that achieves quasi-orthogonal textual tuning by Controlling language hyperspherical Energy (CarE) to remove the additional visual encoder for enabling collaborative multi-modal gradients in VioLET-CMG. VioLET-CarE enhances computational efficiency via additive low-rank feature transformation for textual tuning. Experimental results demonstrate that the proposed VioLET framework consistently achieves state-of-the-art performance in new class generalization and few-shot recognition using both ResNet-50 and ViT/B-16 backbones. Remarkably, compared with VioLET-CMG, VioLET-CarE further reduces GPU memory cost by 27% and training times by 36% on average with enhanced generalization ability. [ABSTRACT FROM AUTHOR]
– Name: AbstractSuppliedCopyright
  Label:
  Group: Ab
  Data: <i>Copyright of International Journal of Computer Vision is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract.</i> (Copyright applies to all Abstracts.)
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=egs&AN=192202262
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1007/s11263-025-02690-2
    Languages:
      – Code: eng
        Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 28
        StartPage: 1
    Subjects:
      – SubjectFull: Orthogonalization
        Type: general
      – SubjectFull: Generalization
        Type: general
      – SubjectFull: Machine learning
        Type: general
    Titles:
      – TitleFull: Vision-Language Efficient Tuning for Mitigating Catastrophic Forgetting in Multi-Modal Learning.
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Wang, Yaoming
      – PersonEntity:
          Name:
            NameFull: Liu, Yuchen
      – PersonEntity:
          Name:
            NameFull: Dai, Wenrui
      – PersonEntity:
          Name:
            NameFull: Zhang, Xiaopeng
      – PersonEntity:
          Name:
            NameFull: Li, Han
      – PersonEntity:
          Name:
            NameFull: Li, Chenglin
      – PersonEntity:
          Name:
            NameFull: Zou, Junni
      – PersonEntity:
          Name:
            NameFull: Tian, Qi
      – PersonEntity:
          Name:
            NameFull: Xiong, Hongkai
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 04
              Text: Apr2026
              Type: published
              Y: 2026
          Identifiers:
            – Type: issn-print
              Value: 09205691
          Numbering:
            – Type: volume
              Value: 134
            – Type: issue
              Value: 4
          Titles:
            – TitleFull: International Journal of Computer Vision
              Type: main
ResultId 1