Knowledge Commons of Institute of Automation,CAS
Cross-Architecture Knowledge Distillation | |
Yufan Liu1,2; Jiajiong Cao5; Bing Li1,4; Weiming Hu1,2,3; Jingting Ding5; Liang Li5 | |
2022-12 | |
会议名称 | Proceedings of the Asian Conference on Computer Vision (ACCV) |
会议录名称 | INTERNATIONAL JOURNAL OF COMPUTER VISION |
页码 | 27 |
通讯作者 | Li, Bing(bli@nlpr.ia.ac.cn) |
会议日期 | 2022.12.4-2022.12.8 |
会议地点 | Macau SAR, China |
会议主办者 | National Natural Science Foundation of China ; National Key Research and Development Program of China ; Project of Beijing Science and technology Committee ; Beijing Natural Science Foundation ; Major Projects of Guangdong Education Department for Foundation Research and Applied Research ; Guangdong Provincial University Innovation Team Project ; Youth Innovation Promotion Association, CAS |
出版者 | SPRINGER |
摘要 | Transformer attracts much attention because of its ability to learn global relations and superior performance. In order to achieve higher performance, it is natural to distill complementary knowledge from Transformer to convolutional neural network (CNN). However, most existing knowledge distillation methods only consider homologous-architecture distillation, such as distilling knowledge from CNN to CNN. They may not be suitable when applying to cross-architecture scenarios, such as from Transformer to CNN. To deal with this problem, a novel cross-architecture knowledge distillation method is proposed. Specifically, instead of directly mimicking output/intermediate features of the teacher, partially cross attention projector and group-wise linear projector are introduced to align the student features with the teacher's in two projected feature spaces. And a multi-view robust training scheme is further presented to improve the robustness and stability of the framework. Extensive experiments show that the proposed method outperforms 14 state-of-the-arts on both small-scale and large-scale datasets. |
关键词 | Knowledge distillation Cross architecture Model compression Deep learning |
DOI | 10.1007/s11263-024-02002-0 |
收录类别 | SCI |
资助项目 | National Natural Science Foundation of China ; National Key Research and Development Program of China[2020AAA0105802] ; National Key Research and Development Program of China[2020AAA0105800] ; Project of Beijing Science and technology Committee[Z231100005923046] ; Beijing Natural Science Foundation[L223003] ; Major Projects of Guangdong Education Department for Foundation Research and Applied Research[2017KZDXM081] ; Major Projects of Guangdong Education Department for Foundation Research and Applied Research[2018KZDXM066] ; Guangdong Provincial University Innovation Team Project[2020KCXTD045] ; Youth Innovation Promotion Association, CAS ; [62192785] ; [62372451] ; [62372082] ; [U1936204] ; [62272125] ; [62306312] ; [62036011] ; [62192782] ; [61721004] ; [U2033210] |
语种 | 英语 |
WOS研究方向 | Computer Science |
WOS类目 | Computer Science, Artificial Intelligence |
WOS记录号 | WOS:001164298900001 |
七大方向——子方向分类 | 图像视频处理与分析 |
国重实验室规划方向分类 | 智能计算与学习 |
是否有论文关联数据集需要存交 | 否 |
引用统计 | |
文献类型 | 会议论文 |
条目标识符 | http://ir.ia.ac.cn/handle/173211/51486 |
专题 | 多模态人工智能系统全国重点实验室_视频内容安全 |
通讯作者 | Bing Li |
作者单位 | 1.Institute of Automation Chinese Academy of Sciences 2.School of Artificial Intelligence, University of Chinese Academy of Sciences 3.CAS Center for Excellence in Brain Science and Intelligence Technology 4.PeopleAI, Inc. 5.Ant Financial Service Group |
第一作者单位 | 中国科学院自动化研究所 |
通讯作者单位 | 中国科学院自动化研究所 |
推荐引用方式 GB/T 7714 | Yufan Liu,Jiajiong Cao,Bing Li,et al. Cross-Architecture Knowledge Distillation[C]:SPRINGER,2022:27. |
条目包含的文件 | 下载所有文件 | |||||
文件名称/大小 | 文献类型 | 版本类型 | 开放类型 | 使用许可 | ||
ACCV2022_CrossArchKD(1020KB) | 会议论文 | 开放获取 | CC BY-NC-SA | 浏览 下载 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论