Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Learning to Obstruct Few-Shot Image Classification over Restricted Classes
Amber Yijia Zheng, Chiao-An Yang, Raymond A. Yeh
Read the paperCheck your idea against 137,217 papers
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 LTO backbone (poor initialization) before CE few-shot learning on CIFAR100 | CLIP-RN50 + CE · CIFAR100 (avg over superclasses) | DropRatio@2: 1.85 -> 10.76 | Improved |
| include the FSC learner in the obstruction objective (LTO vs No F) (ablation) | LTO · CIFAR100 CE | DropRatio@2: 5.97 -> 10.76 | Improved |
| include the FSC learner in the obstruction objective (LTO vs No F) on CoOp (ablation) | LTO · CIFAR100 CoOp | DropRatio@2: 4.54 -> 4.80 | No clear effect |
| maximize restricted-class loss only (Only R) vs LTO (ablation) | LTO · SUN397 MetaOptNet 5-shot | DropRatio@2: 18.01 -> 1.17 | Hurt |
| use LTO instead of machine unlearning (SSD) to obstruct | CLIP-RN50 + CE · CIFAR100 first 9 superclasses | DropRatio@2: 1.09 -> 11.99 | Improved |
| give the evaluation few-shot learner 4x more data (ablation) | LTO-obstructed CLIP-RN50 + CE · ImageNet superclass 1 | DropRatio@2: 9.93 -> 2.82 | Hurt |
| train the evaluation few-shot learner 4x more epochs (ablation) | LTO-obstructed CLIP-RN50 + Tip-Adapter · ImageNet superclass 1 | DropRatio@2: 5.92 -> 3.16 | Hurt |
| evaluate with a different few-shot method than the one used in LTO (ablation) | LTO · ImageNet superclass 1 | DropRatio@2: Tip-Adapter LTO: 5.92 on TA, 4.16 on CE, 7.75 on CoOp | Mixed |
| increase SSD training data from 5-shot to 20-shot (ablation) | SSD unlearning baseline · CIFAR100 | DropRatio@2: 0.67 -> 2.22 (CE, superclass 1) | Improved |
| apply LTO to CLIP attribute learning on CelebA | CLIP-RN50 + CE · CelebA attributes | AUROC drop ratio: 8.04 average; Pale_Skin 31.64, Gray_Hair 0.08 | Improved |
Closest papers by meaning
- Rethinking Few-shot Image Classification: A Good Embedding is All You Need? (ECCV 2020 · similarity 0.89)
- Few-Shot Image Classification: Just Use a Library of Pre-Trained Feature Extractors and a Simple Classifier (ICCV 2021 · similarity 0.89)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.89)
- Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference (CVPR 2022 · similarity 0.88)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.88)
- MetaCoCo: A New Few-Shot Classification Benchmark with Spurious Correlation (ICLR 2024 · similarity 0.88)
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.
