Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Few-Shot Image Classification: Just Use a Library of Pre-Trained Feature Extractors and a Simple Classifier
Arkabandhu Chowdhury, Mingchao Jiang, Swarat Chaudhuri, Chris Jermaine
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 |
|---|---|---|---|
| concatenate features of all nine library CNNs instead of using the best single CNN (ablation) | MLP few-shot learner · Aircraft 20-way 5-shot | accuracy: 44.6 -> 49.5 | Improved |
| concatenate features of all nine library CNNs instead of using the best single CNN (ablation) | MLP few-shot learner · Aircraft 40-way 5-shot | accuracy: 35.9 -> 41.2 | Improved |
| concatenate features of all nine library CNNs instead of using the best single CNN (ablation) | MLP few-shot learner · Aircraft 5-way 5-shot | accuracy: 66.0 -> 68.9 | Improved |
| use one MLP on concatenated features instead of a hard ensemble of per-CNN classifiers (ablation) | Full Library · Aircraft 20-way 5-shot | accuracy: 46.6 -> 49.5 | Improved |
| use one MLP on concatenated features instead of a soft ensemble of per-CNN classifiers (ablation) | Full Library · Aircraft 40-way 5-shot | accuracy: 39.0 -> 41.2 | Improved |
| use concatenated features instead of ensembling on 5-way tasks (ablation) | Full Library · FC100 5-way 5-shot | accuracy: 79.1 vs 80.5 | No clear effect |
| use library of nine ILSVRC2012 CNNs instead of a single BiT CNN trained on full ImageNet (ablation) | Full Library · Aircraft 40-way 5-shot | accuracy: 33.0 -> 41.2 | Improved |
| use library of nine ILSVRC2012 CNNs instead of a single BiT CNN trained on full ImageNet (ablation) | Full Library · VGG Flower 5-way 5-shot | accuracy: 97.4 vs 99.7 | Hurt |
| use library of nine ILSVRC2012 CNNs instead of a single BiT CNN (ablation) | Full Library · Traffic Sign 20-way 5-shot | accuracy: 53.5 -> 70.4 | Improved |
| replace meta-learning/transfer baselines with a library-based learner | best library-based learner · Aircraft 5-way 5-shot | accuracy: 47.6 -> 66.0 | Improved |
| replace SUR with a library-based learner | best library-based learner · Omniglot 5-way 5-shot | accuracy: 98.7 vs 96.7 | Hurt |
| train on 1-shot vs full data without regularization: compare learned weights | linear classifier on 13,984 features · Traffic Sign 40-way | Pearson correlation: 0.95 on Aircraft vs 0.18 on Traffic Sign | Mixed |
Closest papers by meaning
- Rethinking Few-shot Image Classification: A Good Embedding is All You Need? (ECCV 2020 · similarity 0.91)
- Selecting Relevant Features from a Multi-domain Representation for Few-shot Classification (ECCV 2020 · similarity 0.90)
- Few-Shot Image Recognition by Predicting Parameters From Activations (CVPR 2018 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.90)
- Unraveling Meta-Learning: Understanding Feature Representations for Few-Shot Tasks (ICML 2020 · similarity 0.90)
- LibFewShot: A Comprehensive Library for Few-Shot Learning (TPAMI 2023 · similarity 0.90)
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.
