Research map / Architectures and efficiency / RNNs and normalization
Towards a Smaller Student: Capacity Dynamic Distillation for Efficient Image Retrieval
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 |
|---|---|---|---|
| Add DGC module (decoupling KD and pruning on separate 1x1 layer) (ablation) | CDD student ResNet101 · VeRi776 | mAP (%): +18.22 mAP over CDD w/o DGC | Helped |
| Add DGC module (ablation) | CDD student ResNet101 · In-Shop | mAP (%): 71.87 to 81.38 mAP | Helped |
| Add RGGR gradient resetting on top of CDD (ablation) | CDD student ResNet101 · VeRi776 | FLOPs (G) / MP (M): MP 18.59M to 14.30M; FLOPs 5.77G to 4.46G, mAP 80.36 to 80.67 | Helped |
| Add RGGR on top of CDD (ablation) | CDD student ResNet101 · In-Shop | FLOPs (G): 6.31G to 4.68G (mAP 81.38 to 81.28) | Helped |
| Use end-to-end KD+pruning (CDD+RGGR) instead of two-stage self-KD followed by ResRep pruning (ablation) | ResNet101 student on MSMT17 · MSMT17 | mAP (%): +4.81 mAP (56.17 to 60.98) | Helped |
| Increase p (fraction of channels zeroed by RGGR) from 40% to 70% (ablation) | CDD+RGGR · In-Shop | mAP (%) / FLOPs (G): mAP down 0.27; FLOPs 4.9G to 4.3G | Mixed |
| Use a heavy same-depth ResNet101 student (instead of lightweight student) with KD and pruning (ablation) | CDD on VeRi776 without RGGR/DGC · VeRi776 | mAP (%): 62.14 mAP vs teacher 80.50 | Hurt |
| Use CDD+RGGR (heavy-to-slim ResNet101 student) instead of lightweight ResNet34 KD students | ResNet101 student · VeRi776 | mAP (%): 80.67 vs 78.56 (RKD, best KD baseline) | Helped |
| Use CDD+RGGR instead of ResRep structured pruning | ResNet101 student · In-Shop | mAP (%): 81.28 vs 79.86 | Helped |
| Use CDD+RGGR with ResNet50 teacher/student on VeRi776 instead of VKD self-distillation | ResNet50 student · VeRi776 | mAP (%): 78.75 vs 79.17 (VKD) | Hurt |
Closest papers by meaning
- Let All Be Whitened: Multi-Teacher Distillation for Efficient Visual Retrieval (AAAI 2024 · similarity 0.89)
- Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation (NeurIPS 2024 · similarity 0.87)
- Dynamic Contrastive Knowledge Distillation for Efficient Image Restoration (AAAI 2025 · similarity 0.87)
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.86)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.86)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.86)
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.