Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Associative Alignment for Few-shot Image Classification
Arman Afrasiyabi, Jean-François Lalonde, Christian Gagné
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 centroid associative alignment of related base classes | arcmax baseline, ResNet-18 · mini-ImageNet 5-shot | accuracy: 76.62 -> 80.35 | Improved |
| add adversarial associative alignment of related base classes | arcmax baseline, ResNet-18 · mini-ImageNet 5-shot | accuracy: 76.62 -> 77.92 | Improved |
| add centroid associative alignment | arcmax baseline, Conv4 · mini-ImageNet 5-shot | accuracy: 69.07 -> 71.45 | Improved |
| use adversarial instead of centroid alignment | alignment, Conv4 · CUB 1-shot | accuracy: 62.71 -> 63.30 | Improved |
| add adversarial alignment in cross-domain transfer | arcmax baseline, ResNet-18 · mini-ImageNet to CUB 1-shot | accuracy: 45.60 -> 44.37 | Hurt |
| add centroid alignment in cross-domain transfer | arcmax baseline, ResNet-18 · mini-ImageNet to CUB 5-shot | accuracy: 64.93 -> 70.37 | Improved |
| add alignment loss (centroid) to training on novel plus related base classes (ablation) | related-base fine-tuning, ResNet-18 · mini-ImageNet 1-shot | accuracy: 55.94 -> 59.88 | Improved |
| train on related base classes without alignment loss (ablation) | arcmax baseline, Conv4 · mini-ImageNet 1-shot | accuracy: 51.90 -> 52.72 | No clear effect |
| increase number of related base classes B from 1 to 10 (centroid) (ablation) | centroid alignment, ResNet-18 · mini-ImageNet 5-shot | accuracy: 77.54 -> 80.23 | Improved |
| increase number of related base classes B from 1 to 10 (adversarial) (ablation) | adversarial alignment, ResNet-18 · mini-ImageNet 5-shot | accuracy: 79.34 -> 77.92 | Hurt |
| use selected related base classes instead of random ones (ablation) | centroid alignment, ResNet-18 · mini-ImageNet 5-shot | accuracy: 76.19 -> 80.35 | Improved |
| use arcmax with early stopping instead of cosmax (ablation) | baseline, Conv4 · mini-ImageNet 5-shot | accuracy: 68.47 -> 69.07 | Improved |
Closest papers by meaning
- Improved Few-Shot Visual Classification (CVPR 2020 · similarity 0.87)
- Few-Shot Image Recognition by Predicting Parameters From Activations (CVPR 2018 · similarity 0.87)
- FD-Align: Feature Discrimination Alignment for Fine-tuning Pre-Trained Models in Few-Shot Learning (NeurIPS 2023 · similarity 0.87)
- iFAN: Image-Instance Full Alignment Networks for Adaptive Object Detection (AAAI 2020 · similarity 0.86)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.86)
- Adversarially Robust Few-shot Learning via Parameter Co-distillation of Similarity and Class Concept Learners (CVPR 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.
