[arXiv] Group Distributionally Robust Dataset Distillation with Risk Minimization [PDF] [Code]

Saeed Vahidian*, Mingyu Wang*, Jianyang Gu*, Vyacheslav Kungurtsev, Wei Jiang and Yiran Chen.

• Robust Dataset Distillation combines clustering with the minimization of a risk measure on the loss.
• Robust Dataset Distillation substantially enhances the robustness across domain-shifted testing sets.

[CVPR 2024] Efficient Dataset Distillation via Minimax Diffusion [PDF] [Code]

Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You and Yiran Chen.

• Minimax Diffusion only requires 1 hour to finish the distillation for a 10-class subset of full-sized ImageNet.
• Minimax Diffusion generates representative and diverse images, yielding state-of-the-art validation performance.

[AAAI 2024] Summarizing Stream Data for Memory-Restricted Online Continual Learning [PDF] [Code]

Jianyang Gu, Kai Wang, Wei Jiang, and Yang You.

• SSD summarizes the informtion flow in online continual learning into informative samples.
• SSD significantly improves the replay effects, especially for circumstances with restricted memory.

[ICCV 2023] Dataset Quantization [PDF] [Code]

Daquan Zhou*, Kai Wang*, Jianyang Gu*, Dongze Lian, Xiangyu Peng, Yifan Zhang, Yang You, and Jiashi Feng.

• DQ targets at compressing large-scale datasets into subsets with lossless validation performance.
• DQ only requires 60% ImageNet and 20% Alpaca's instruction tuning data to achieve lossless training.

[ICCV 2023] DREAM: Efficient Dataset Distillation by Representative Matching [PDF] [Code]

Yanqing Liu*, Jianyang Gu*, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You.

• DREAM aims to improve the training efficiency by only calculating matching metrics with representative samples.
• DREAM only requires less than 20% of original iterations to achieve baseline performance.

[arXiv] DiM: Distilling Dataset into Generative Model [PDF] [Code]

Kai Wang*, Jianyang Gu*, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You.

• DiM opens up a new generative scheme for dataset distillation by condensing the information into models.
• DiM generates diverse images on the fly, providing outstanding distillation performance with small extra training consumption.

[arXiv] Color Prompting for Data-Free Continual Unsupervised Domain Adaptive Person Re-Identification [PDF] [Code]

Jianyang Gu, Hao Luo, Kai Wang, Wei Jiang, Yang You, and Jian Zhao.

• CoP learns the color distribution conditioned on the tasks and image contents during the continual learning process.
• CoP provides substantial improvements on anti-forgetting and generalization without storing images of previous tasks.

[CVPR 2023] MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID [PDF] [Code]

Jianyang Gu, Kai Wang, Hao Luo, Chen Chen, Wei Jiang, Yuqiang Fang, Shanghang Zhang, Yang You, and Jian Zhao.

• MSINet searches the optimal interation between multi-scale branches with a twins contrastive mechanism.
• MSINet yeilds better performance on both supervised and unsupervised tasks with limited parameter size.

[IEEE TIFS] Multi-View Evolutionary Training for Unsupervised Domain Adaptive Person Re-Identification [PDF]

Jianyang Gu, Weihua Chen, Hao Luo, Fan Wang, Hao Li, Wei Jiang, and Weijie Mao.

• MET effectively reduces the clustering noise from the dimensions of snapshot quality and temporal consistency.
• MET significantly improves the training robustness by high-quality clustering results, leading to better model performance.