Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
DiffKendall: A Novel Approach for Few-Shot Learning with Differentiable Kendall's Rank Correlation
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 |
|---|---|---|---|
| replace cosine similarity with Kendall rank correlation at test time only (ablation) | CE-pretrained ResNet-12 · Traffic Signs 5-way 1-shot | accuracy (%): 55.54 -> 60.84 | Helped |
| replace cosine with Kendall at test time only (ablation) | CE-pretrained WRN-28-10 · Traffic Signs 5-way 1-shot | accuracy (%): 47.01 -> 51.39 | Helped |
| replace cosine with Kendall at test time only (ablation) | S2M2 WRN-28-10 · mini-test 5-way 1-shot | accuracy (%): 64.52 -> 63.97 | Hurt |
| replace cosine with Kendall at test time only (ablation) | CE-pretrained ResNet-18 · mini-test 5-way 1-shot | accuracy (%): 62.92 -> 62.83 | No clear effect |
| replace cosine with Kendall at test time vs CIM feature transformation | CE-pretrained ResNet-12 · mini-test 5-way 1-shot | accuracy (%): 60.4 -> 63.3 | Helped |
| replace cosine with differentiable Kendall in episodic training of Meta-Baseline | Meta-Baseline ResNet-12 · mini-ImageNet 5-way 1-shot | accuracy (%): 63.17 -> 65.56 | Helped |
| replace cosine with differentiable Kendall in episodic training of Meta-Baseline | Meta-Baseline ResNet-12 · tiered-ImageNet 5-way 5-shot | accuracy (%): 83.74 -> 85.31 | Helped |
| use differentiable Kendall in episodic training in addition to test-time Kendall (ablation) | Meta-Baseline ResNet-12 · 5-way 1-shot on varied domains | accuracy (%): 1%-2% improvement | Helped |
| use only small-valued channels (mask large ones), Kendall vs cosine (ablation) | mini-ImageNet features · mini-test 5-way 1-shot | accuracy (%): approximately 9% improvement | Helped |
| mask small-valued channels, Kendall vs cosine (ablation) | mini-ImageNet features · mini-test 5-way 1-shot | accuracy (%) | No clear effect |
| sample 5n channel pairs instead of all pairs for Kendall (ablation) | Kendall test-time similarity · mini-ImageNet | accuracy (%) | No clear effect |
| vary alpha of smooth approximation (too large or too small) (ablation) | DiffKendall · mini-ImageNet | accuracy (%) | Hurt |
Closest papers by meaning
- Finding Task-Relevant Features for Few-Shot Learning by Category Traversal (CVPR 2019 · similarity 0.90)
- Exploring Complementary Strengths of Invariant and Equivariant Representations for Few-Shot Learning (CVPR 2021 · similarity 0.90)
- Adaptive Poincaré Point to Set Distance for Few-Shot Classification (AAAI 2022 · similarity 0.88)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.88)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.87)
- Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · 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.