Research map / Architectures and efficiency / Attention and transformers
Attribute Surrogates Learning and Spectral Tokens Pooling in Transformers for Few-Shot Learning
Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, Wenqiang Zhang
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add class and patch attribute-surrogate losses to DINO training (ablation) | DINO ViT-S · miniImageNet | 5-way 1-shot accuracy: 61.57 ± 0.16 -> 71.27 ± 0.17 | Improved |
| Replace surrogate losses with cross-entropy supervision (ablation) | DINO ViT-S · miniImageNet | 5-way 1-shot accuracy: 71.27 ± 0.17 -> 66.81 ± 0.17 | Hurt |
| Add the class-token surrogate loss (ablation) | DINO with patch surrogate loss · miniImageNet | 5-way 1-shot accuracy: 63.17 ± 0.16 -> 71.27 ± 0.17 | Improved |
| Add the patch-token surrogate loss (ablation) | DINO with class surrogate loss · miniImageNet | 5-way 1-shot accuracy: 68.95 ± 0.17 -> 71.27 ± 0.17 | Improved |
| Set patch-surrogate loss weight β to 1 instead of 0.1 (ablation) | HCTransformers first student · miniImageNet | 5-way 1-shot accuracy: 71.27 ± 0.17 -> 61.45 ± 0.16 | Hurt |
| Set patch-surrogate loss weight β to 0.01 instead of 0.1 (ablation) | HCTransformers first student · miniImageNet | 5-way 1-shot accuracy: 71.27 ± 0.17 -> 70.40 ± 0.16 | Hurt |
| Use 224²/8 rather than 56²/2 input/patch settings (ablation) | HCTransformers first student · CIFAR-FS | 5-way 1-shot accuracy: 71.99 ± 0.17 -> 73.28 ± 0.16 | Improved |
| Use 224²/8 rather than 112²/4 input/patch settings (ablation) | HCTransformers first student · CIFAR-FS | 5-way 1-shot accuracy: 72.80 ± 0.17 -> 73.28 ± 0.16 | Improved |
| Train the latter transformer sets one-by-one rather than end-to-end (ablation) | HCTransformers · miniImageNet | stage 2 1-shot accuracy: 74.74 ± 0.17 -> 74.40 ± 0.17 | Hurt |
| Train the latter transformer sets one-by-one rather than end-to-end (ablation) | HCTransformers · miniImageNet | stage 3 1-shot accuracy: 72.66 ± 0.18 -> 73.01 ± 0.18 | Improved |
| Supervise latter-stage patch tokens in addition to class tokens (ablation) | HCTransformers later stages · miniImageNet | stage 2 1-shot accuracy: 74.74 ± 0.17 -> 71.25 ± 0.17 | Hurt |
| Supervise class surrogate using global and local views instead of global views only (ablation) | HCTransformers first student · miniImageNet | 5-way 1-shot accuracy: 71.27 ± 0.17 -> 69.37 ± 0.16 | Hurt |
Closest papers by meaning
- Scalable Vision Transformers With Hierarchical Pooling (ICCV 2021 · similarity 0.87)
- TransVLAD: Focusing on Locally Aggregated Descriptors for Few-Shot Learning (ECCV 2022 · similarity 0.86)
- Supervised Masked Knowledge Distillation for Few-Shot Transformers (CVPR 2023 · similarity 0.86)
- A Theoretical Understanding of Shallow Vision Transformers: Learning, Generalization, and Sample Complexity (ICLR 2023 · similarity 0.86)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.86)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
