Research map / Label-efficient and robust learning / Self-supervised and contrastive
Slimmed Asymmetrical Contrastive Learning and Cross Distillation for Lightweight Model Training
Jian Meng, Li Yang, Kyungmin Lee, Jinwoo Shin, Deliang Fan, Jae-sun Seo
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 cross-distillation (XD) to Barlow-Twins-style loss on 1x model, no teacher | MobileNet-V3 · ImageNet-1K linear eval | top-1 acc (%): 56.70 -> 59.42 | Improved |
| Add asymmetrical slimmed contrastive learning (SACL) on top of XD (ablation) | EfficientNet-B0 · ImageNet-1K linear eval | top-1 acc (%): 65.32 | Improved |
| Increase asymmetry from 2x-1x to 6x-1x (ablation) | SACL+XD ResNet-20 · CIFAR-10 linear eval | acc (%): 82.31 -> 86.81 | Improved |
| Train 100 vs 200 epochs with XD only (ablation) | MobileNet-V3 · ImageNet-1K | top-1 acc (%): 57.16 -> 59.42 | Improved |
| Extend training from 100 to 300 epochs (ablation) | MobileNet-V1 XD · ImageNet-1K | top-1 acc (%): 58.25 | Improved |
| Extend training to 300 epochs (ablation) | MobileNet-V1 SACL+XD · ImageNet-1K | top-1 acc (%): 60.63 | Improved |
| Replace inner-decorrelation distillation loss with negative-log loss (ablation) | XD MobileNet-V1 · ImageNet-100 | acc (%): 80.30 -> 79.63 | Hurt |
| Use XD instead of Barlow Twins | MobileNet-V1 · ImageNet-100 | acc (%): 78.40 -> 80.30 | Improved |
| Use XD instead of Barlow Twins | ResNet-50 · ImageNet-1K 300 epochs | top-1 acc (%): 70.7 -> 71.1 | Improved |
| Use XD instead of DINO | ViT-Tiny · ImageNet-100 | acc (%): 63.04 -> 63.92 | Improved |
| Vary alpha weight of inner decorrelation loss (ablation) | XD MobileNet-V1 · ImageNet-1K | top-1 acc (%): best at alpha=0.9; within ± 1% | No clear effect |
| Minimize only the infinitesimal inner correlation as LCD (ablation) | SACL · training | stability/accuracy | 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.
- Barlow Twins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 4 of 4 shared results
- VICReg — VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning (ICLR 2022) · 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
- Slimmable Networks for Contrastive Self-supervised Learning (IJCV 2024 · similarity 0.90)
- DisCo: Remedying Self-Supervised Learning on Lightweight Models with Distilled Contrastive Learning (ECCV 2022 · similarity 0.89)
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.88)
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.88)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · 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.
