Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Low-Shot Learning With Large-Scale Diffusion
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 |
|---|---|---|---|
| Use in-domain unlabeled images for label diffusion | Low-shot ImageNet classification · ImageNet val, n=1 | top-5 accuracy: 60.4±0.78 -> 69.7±0.86 | Helped |
| Increase the number of out-of-domain YFCC images from F1M to F100M | Diffusion · ImageNet val, n=2 | top-5 accuracy: 66.8±0.71 -> 69.5±0.60 | Helped |
| Fuse diffusion scores with logistic regression scores | Low-shot ImageNet classification · ImageNet val, out-of-domain F100M, n=2 | top-5 accuracy: 69.5±0.60 -> 71.1±0.82 | Helped |
| Use in-domain rather than out-of-domain background images | Diffusion · ImageNet val, n=2 | top-5 accuracy: 69.5±0.60 -> 75.4±0.64 | Helped |
| Omit normalization of propagated labels (ablation) | Diffusion · ImageNet val, group 1, n=2, no background | top-5 accuracy: 62.7±0.68 -> 40.6±0.18 | Hurt |
| Use Sinkhorn projection instead of column-wise normalization (ablation) | Diffusion · ImageNet val, group 1, n=2, F1M | top-5 accuracy: 65.4±0.55 -> 56.8±0.50 | Hurt |
| Weight edges by meaningful-neighbor criterion (ablation) | Diffusion · ImageNet val, group 1, n=2, F1M | top-5 accuracy: 65.4±0.55 -> 40.0±0.20 | Hurt |
| Use Gaussian edge weights (ablation) | Diffusion · ImageNet val, group 1, n=2, F1M | top-5 accuracy: 65.4±0.55 -> 65.4±0.58 | No clear effect |
| Apply nonlinear Γr transform to diffusion updates (ablation) | Diffusion · ImageNet val, group 1, n=2, F1M | top-5 accuracy: 65.4±0.55 -> 65.4±0.55 | No clear effect |
| Apply class-frequency prior (ablation) | Diffusion · ImageNet val, group 1, n=2, F1M | top-5 accuracy: 65.4±0.55 -> 65.4±0.60 | No clear effect |
| Use in-domain diffusion instead of logistic regression | Low-shot ImageNet classification · ImageNet val, n=20 | top-5 accuracy: 86.0±0.15 -> 83.6±0.12 | Hurt |
Closest papers by meaning
- Few-Shot Learning with Graph Neural Networks (ICLR 2018 · similarity 0.88)
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.88)
- Meta-Learning for Semi-Supervised Few-Shot Classification (ICLR 2018 · similarity 0.88)
- Label Efficient Semi-Supervised Learning via Graph Filtering (CVPR 2019 · similarity 0.88)
- Transductive Few-Shot Learning With Prototype-Based Label Propagation by Iterative Graph Refinement (CVPR 2023 · similarity 0.87)
- Semi-Supervised Learning for Few-Shot Image-to-Image Translation (CVPR 2020 · similarity 0.87)
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.