Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Diversity With Cooperation: Ensemble Methods for Few-Shot Classification
Nikita Dvornik, Cordelia Schmid, Julien Mairal
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 cooperation penalty (symmetrized KL on non-ground-truth probabilities) to a small ensemble of 5 networks | independent ensemble · miniImageNet 5-shot | accuracy (%): 80.02 -> 80.60 | Improved |
| add diversity penalty to a large ensemble of 20 networks | independent ensemble · miniImageNet 5-shot | accuracy (%): 80.57 -> 81.18 | Improved |
| add robust strategy (cooperation plus network dropping, dropout, different augmentations) to 20 networks | independent ensemble · miniImageNet 1-shot | accuracy (%): 62.06 -> 63.95 | Improved |
| add cooperation penalty to a 20-network ensemble | independent ensemble · miniImageNet 5-shot | accuracy (%): 80.57 -> 80.80 | No clear effect |
| distill 20-network robust ensemble into one network using extra unlabeled COCO images | robust ensemble · miniImageNet 5-shot | accuracy (%): 81.59 -> 81.19 | Hurt |
| add extra unannotated images to distillation (ablation) | robust-dist, 20 networks · miniImageNet 5-shot | accuracy (%): 80.63 -> 81.19 | Improved |
| add extra unannotated COCO images to distillation on CUB (ablation) | robust-dist, 20 networks · CUB 5-shot | accuracy (%): 83.57 -> 83.21 | Hurt |
| use conditional non-ground-truth probabilities instead of full softmax probabilities in cooperation penalty (ablation) | cooperation ensemble, 5 networks · miniImageNet 5-shot | accuracy (%): 79.79 -> 80.60 | Improved |
| use symmetrized KL instead of L2 for cooperation (ablation) | cooperation ensemble, 5 networks · miniImageNet 5-shot | accuracy (%): 80.14 -> 80.72 | Improved |
| use negative KL instead of cosine for diversity (ablation) | diversity ensemble, 5 networks · miniImageNet 5-shot | accuracy (%): 80.82 -> 79.81 | Hurt |
| use cosine instead of L2 for diversity (ablation) | diversity ensemble, 5 networks · miniImageNet 5-shot | accuracy (%): 80.54 -> 80.82 | No clear effect |
| use learned weights for class prototypes instead of mean centroid (ablation) | mean centroid classifier · miniImageNet 5-shot | accuracy (%): 77.28 -> 77.01 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TADAM — TADAM: Task dependent adaptive metric for improved few-shot learning (NeurIPS 2018) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Theoretical Analysis of the Number of Shots in Few-Shot Learning (ICLR 2020 · similarity 0.90)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.90)
- Diversity Transfer Network for Few-Shot Learning (AAAI 2020 · similarity 0.90)
- Meta-Learning With Differentiable Convex Optimization (CVPR 2019 · similarity 0.90)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.89)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 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.
