Research map / Label-efficient and robust learning / Knowledge distillation
Grouped Knowledge Distillation for Deep Face Recognition
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 |
|---|---|---|---|
| omit Secondary-KD while retaining Primary-KD and Binary-KD (ablation) | classical KD for face recognition · CFP-FP | ACC: 91.71 -> 94.35 | Helped |
| add Binary-KD to Primary-KD (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 94.08 -> 94.35 | Helped |
| reduce cumulative probability threshold from 1 to 0.93 (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 91.71 -> 94.35 | Helped |
| set cumulative probability threshold to 0.91 rather than 0.93 (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 94.35 -> 94.00 | Hurt |
| increase Primary-KD weight from λ1=2 to λ1=8 (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 94.07 -> 94.35 | Helped |
| increase Binary-KD weight from λ2=0 to λ2=1 (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 94.08 -> 94.35 | Helped |
| vary Binary-KD weight from λ2=1 to λ2=2 (ablation) | Grouped Knowledge Distillation · CFP-FP | ACC: 94.35 -> 94.23 | No clear effect |
| apply GKD instead of EKD | MobileFaceNet · IJB-C | TPR@FPR=1e-4: 90.48 -> 94.34 | Helped |
| apply GKD instead of EKD | MobileFaceNet · IJB-C | TPR@FPR=1e-5: 84.00 -> 91.01 | Helped |
| apply GKD instead of EKD | MobileFaceNet · CPLFW | ACC: 89.35 -> 90.86 | Helped |
| apply GKD rather than train from scratch | MobileNetV3-L · CALFW M-M | ACC: 92.65 -> 94.41 | Helped |
Closest papers by meaning
- Debiased Distillation for Consistency Regularization (AAAI 2025 · similarity 0.86)
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.86)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.86)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.86)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.85)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 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.