Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Label Hallucination for Few-Shot Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 soft pseudo-label distillation on the full base dataset during finetuning (ablation) | few-shot transfer with ResNet-12 · miniImageNet 5-way 5-shot | accuracy: 80.03 -> 85.60 | Helped |
| Add soft pseudo-label distillation on the full base dataset during finetuning (ablation) | few-shot transfer with ResNet-12 · CIFAR-FS 5-way 5-shot | accuracy: 85.7 -> 89.5 | Helped |
| Add soft pseudo-label distillation on the full base dataset during finetuning (ablation) | few-shot transfer with ResNet-12 · FC100 5-way 5-shot | accuracy: 61.9 -> 67.2 | Helped |
| Use soft pseudo-labels instead of one-hot hard labels (ablation) | Label-Halluc · miniImageNet 5-way 5-shot | accuracy: 80.68 -> 85.60 | Helped |
| Use soft pseudo-labels instead of one-hot hard labels (ablation) | Label-Halluc · CIFAR-FS 5-way 5-shot | accuracy: 85.3 -> 89.5 | Helped |
| Use soft pseudo-labels instead of one-hot hard labels (ablation) | Label-Halluc · FC100 5-way 5-shot | accuracy: 62.4 -> 67.2 | Helped |
| Apply hallucinated-label gradients to the backbone as well as the classifier (ablation) | Label-Halluc · miniImageNet 5-way 5-shot | accuracy: 81.53 -> 85.60 | Helped |
| Apply hallucinated-label gradients to the backbone rather than classifier only (ablation) | Label-Halluc · miniImageNet 5-way 5-shot | accuracy: 81.53 -> 84.36 | Helped |
| Finetune the entire network using only the few-shot support set rather than keeping the backbone frozen (ablation) | few-shot transfer with ResNet-12 · miniImageNet 5-way 5-shot | accuracy: 83.18 -> 80.03 | Hurt |
| Add Label-Halluc finetuning to IER-distill-pretrained features | Label-Halluc (IER-distill pretrained) · miniImageNet 5-way 5-shot | accuracy: 84.50 -> 86.54 | Helped |
| Add Label-Halluc finetuning to IER-distill-pretrained features | Label-Halluc (IER-distill pretrained) · tieredImageNet 5-way 5-shot | accuracy: 86.57 ± 0.81 -> 87.68 ± 0.83 | Helped |
| Use soft pseudo-label distillation rather than AssoAlign softMax with centroid alignment | few-shot transfer with ResNet-12 · FC100 5-way 5-shot | accuracy: 63.6 -> 67.2 | Helped |
Closest papers by meaning
- Meta-Learning for Semi-Supervised Few-Shot Classification (ICLR 2018 · similarity 0.89)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.88)
- Hallucination Improves Few-Shot Object Detection (CVPR 2021 · similarity 0.88)
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.88)
- Dynamic Distillation Network for Cross-Domain Few-Shot Recognition with Unlabeled Data (NeurIPS 2021 · similarity 0.88)
- Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · similarity 0.86)
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.