Research map / Label-efficient and robust learning / Knowledge distillation
Simultaneous Similarity-based Self-Distillation for Deep Metric Learning
Karsten Roth, Timo Milbich, Bjorn Ommer, Joseph Paul Cohen, Marzyeh Ghassemi
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 |
|---|---|---|---|
| Add S2SD to DML objectives | Margin, R-Margin, and Multisimilarity · CUB200-2011, CARS196, and SOP | Recall@1: up to $7\%$ | Improved |
| Add dual similarity self-distillation from one higher-dimensional branch (ablation) | Margin DML · CUB200-2011 | Recall@1: $63.09\pm0.46$ -> $65.11\pm0.18$ | Improved |
| Add multiscale higher-dimensional target branches instead of a single target branch (ablation) | Margin DML · CUB200-2011 | Recall@1: $65.11\pm0.18$ -> $66.13\pm0.34$ | Improved |
| Add feature-space distillation to MSD (ablation) | Margin DML · CUB200-2011 | Recall@1: $66.13\pm0.34$ -> $67.58\pm0.32$ | Improved |
| Add multiple pooling operations in auxiliary branches (ablation) | Margin + MSDF · SOP | Recall@1: $79.63\pm0.14$ -> $78.82\pm0.09$ | Hurt |
| Distill each target space directly to the base space rather than using nested or chained distillation (ablation) | Multiscale self-distillation · DML benchmarks | Recall@1 | Improved |
| Use similarity-vector KL distillation rather than full-matrix/mean KL or cosine/euclidean matching (ablation) | S2SD · DML benchmarks | Recall@1 | Improved |
| Use a two-layer MLP for auxiliary branches rather than a linear mapping (ablation) | S2SD · DML benchmarks | Recall@1 | Improved |
| Increase auxiliary branch depth beyond two-layer MLPs (ablation) | S2SD · DML benchmarks | Recall@1 | Hurt |
| Train auxiliary spaces jointly without similarity distillation (ablation) | S2SD · CUB200-2011 | Recall@1: $62.70\pm0.53$ -> $62.93\pm0.22$ | No clear effect |
| Detach the low-dimensional reference embedding while training in high dimension (ablation) | S2SD · CUB200-2011 | Recall@1: $62.70\pm0.53$ -> $63.08\pm0.42$ | No clear effect |
| Replace simultaneous distillation with standard two-stage distillation (ablation) | DML student, d=128 · CUB200-2011 | Recall@1: $65.57\pm0.26$ -> $63.89\pm0.14$ | Hurt |
Closest papers by meaning
- Revisiting Training Strategies and Generalization Performance in Deep Metric Learning (ICML 2020 · similarity 0.90)
- Improving Deep Metric Learning by Divide and Conquer (TPAMI 2021 · similarity 0.88)
- Deep Compositional Metric Learning (CVPR 2021 · similarity 0.88)
- Deep Disentangled Metric Learning (AAAI 2025 · similarity 0.86)
- Towards Improved Proxy-Based Deep Metric Learning via Data-Augmented Domain Adaptation (AAAI 2024 · similarity 0.86)
- Going Beyond Feature Similarity: Effective Dataset distillation based on Class-aware Conditional Mutual Information (ICLR 2025 · 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.
