Research map / Label-efficient and robust learning / Recognition and long-tail
Debugging and Explaining Metric Learning Approaches: An Influence Function Based Perspective
Ruofan Liu, Yun Lin, XIANGLIN YANG, Jin Song Dong
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 |
|---|---|---|---|
| replace the original influence function with the empirical influence function (EIF) | influence-based debugging of DML · CUB200 individual confusion pairs | average distance improvement: 2.29E-02 at top-30 pairs, p-value 3.47E-12 | Improved |
| replace IF with EIF | influence-based debugging of DML · CARS196 individual confusion pairs | average distance improvement: 1.47E-02 at top-30 pairs, p-value 8.16E-06 | Improved |
| replace IF with EIF | influence-based debugging of DML · InShop individual confusion pairs | average distance improvement: 1.52E-02 at top-30 pairs, p-value 1.96E-05 | Improved |
| replace IF with EIF on groups of confusion pairs | influence-based debugging of DML · InShop group confusion | significance of improvement | Mixed |
| use EIF instead of IF for mislabelled sample detection | influence-based debugging of DML · CUB200 | runtime (s): 375.56 ± 68.06 -> 260.20 ± 29.28 | Improved |
| use EIF instead of IF | influence-based debugging of DML · CARS196 | runtime (s): 588.11 ± 23.23 -> 453.61 ± 28.76 | Improved |
| use EIF instead of IF | influence-based debugging of DML · InShop | runtime (s): 2454.11 ± 9.55 -> 1306.43 ± 27.31 | Improved |
| raise mislabelling ratio from 1% to 10% | EIF relabelling on ProxyNCA++ · CUB200 | relabelling recommendation performance: 75.68% -> 83.28% | Improved |
| raise mislabelling ratio from 1% to 10% | EIF relabelling on SoftTriple · InShop | relabelling recommendation performance: 90.00% -> 86.27% | Hurt |
| raise mislabelling ratio from 1% to 10% | EIF relabelling on SoftTriple · CARS196 | relabelling recommendation performance: 79.49% -> 89.95% | Mixed |
| vary number of sampled parameters N_theta in EIF | EIF · CUB200, CARS196, InShop | ratio of mislabelled samples detected | No clear effect |
| rank samples by influence magnitude, keeping only small-magnitude scores | EIF noisy-sample detection · 3 DML datasets | ratio of mislabelled samples detected | Hurt |
Closest papers by meaning
- Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.88)
- Influence Functions in Deep Learning Are Fragile (ICLR 2021 · similarity 0.86)
- It Takes Two to Tango: Mixup for Deep Metric Learning (ICLR 2022 · similarity 0.86)
- Improving Deep Metric Learning by Divide and Conquer (TPAMI 2021 · similarity 0.86)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.85)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.84)
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.
