Research map / Label-efficient and robust learning / Knowledge distillation
Simple Unsupervised Knowledge Distillation With Space Similarity
Aditya Singh, Haohan Wang
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 space similarity to cosine similarity loss (ResNet-18 student) (ablation) | CoSS · ImageNet-1K linear eval | top-1 accuracy: 61.85 -> 62.35 | Improved |
| add cosine similarity to space similarity loss (EfficientNet-B0 student) (ablation) | CoSS · ImageNet-1K linear eval | top-1 accuracy: 66.76 -> 67.36 | Improved |
| add k=15 nearest-neighbour sampling to batches (ablation) | CoSS · ImageNet-1K linear eval (R18) | top-1 accuracy: 61.89 -> 62.35 | Improved |
| weigh space similarity with lambda 0.5 or 1.0 rather than 0 (ablation) | CoSS · ImageNet-1K linear eval (R18) | top-1 accuracy: 61.85 -> 62.35 | Improved |
| lambda 0.5 vs 1.0 (ablation) | CoSS · ImageNet-1K linear eval (R18) | top-1 accuracy: 62.41 -> 62.35 | No clear effect |
| use CoSS instead of DINO-style soft-label distillation (ablation) | unsupervised distillation · ImageNet-1K linear eval (R18) | top-1 accuracy: 59.88 -> 62.35 | Improved |
| use space similarity instead of BatchNorm-based embedding matching (ablation) | CIFAR-100 distillation · CIFAR-100 ResNet8x4/32x4 | top-1 accuracy: 72.22 -> 73.90 | Improved |
| add space similarity to SEED objective (ablation) | SEED · ImageNet-1K linear eval | top-1 accuracy: 57.60 -> 58.30 | Improved |
| lambda choice on CIFAR-100 ResNet20/56 (ablation) | CoSS · CIFAR-100 | top-1 accuracy: ResNet8x4: 72.05 -> 73.91; ResNet20/56: 70.58 -> 71.34 (best at 0.25), 70.68 at 1 | Mixed |
| CoSS vs PCD on dense prediction | CoSS · MS-COCO and VOC 07+12 | AP: COCO AP bbox 32.2 -> 33.9; VOC AP50 79.4 -> 78.7 | Mixed |
| CoSS vs DisCo on STL-10 transfer | CoSS · STL-10 transfer (ResNet-18) | top-1 accuracy: 95.04 -> 94.11 | Hurt |
| CoSS vs BINGO on ImageNet KNN-10 (ResNet-18) | CoSS · ImageNet-1K KNN-10 | KNN-10 accuracy: 54.16 -> 53.78 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DisCo — DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for Deep Neural Networks (CVPR 2021) · better on 4 of 5 shared results
- SEED — SEED: Self-supervised Distillation For Visual Representation (ICLR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Similarity-Preserving Knowledge Distillation (ICCV 2019 · similarity 0.90)
- SLaM: Student-Label Mixing for Distillation with Unlabeled Examples (NeurIPS 2023 · similarity 0.89)
- Knowledge Distillation Meets Open-Set Semi-supervised Learning (IJCV 2024 · similarity 0.89)
- You Don’t Need Aligned Representations: Knowledge Distillation via Random Prototype Spaces (NeurIPS 2026 · similarity 0.89)
- Exploring Inter-Channel Correlation for Diversity-Preserved Knowledge Distillation (ICCV 2021 · similarity 0.88)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.88)
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.
