Research map / Label-efficient and robust learning / Self-supervised and contrastive
DisCo: Remedying Self-Supervised Learning on Lightweight Models with Distilled Contrastive Learning
Yuting Gao, Jia-Xin Zhuang, Shaohui Lin, Hao Cheng, Xing Sun, Ke Li, Chunhua Shen
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 embedding distillation loss to MoCo-V2 (ablation) | MoCo-V2 student · ImageNet linear eval (teacher ResNet-50) | top-1 accuracy: EfficientNet-B0 46.8 -> 65.6, MobileNet-v3 36.2 -> 63.7, ResNet-18 52.2 -> 57.1 | Improved |
| add contrastive loss to distillation loss (ablation) | DisCo (distillation only) · ImageNet linear eval (teacher ResNet-50) | top-1 accuracy: EfficientNet-B0 65.6 -> 65.6, MobileNet-v3 58.9 -> 63.7, ResNet-18 54.5 -> 57.1 | Mixed |
| enlarge MLP hidden dimension from 512 to 2048 (ablation) | DisCo · ImageNet linear eval (teacher ResNet-50) | top-1 accuracy: EfficientNet-B0 62.5 -> 66.5, ResNet-18 57.1 -> 60.6 | Improved |
| keep a hidden layer in the MLP instead of removing it (ablation) | DisCo · ImageNet linear eval (teacher ResNet-50) | top-1 accuracy: EfficientNet-B0 52.5 -> 66.5, ResNet-18 52.5 -> 60.6 | Improved |
| expand hidden layer to fix Distilling Bottleneck on ResNet-18/34 (ablation) | DisCo* · ImageNet linear eval (teacher ResNet-50) | top-1 accuracy: ResNet-18 57.1 -> 60.6, ResNet-34 58.9 -> 62.5 | Improved |
| use ResNet-152 teacher instead of ResNet-50 (ablation) | DisCo ResNet-34 student · ImageNet linear eval | top-1 accuracy: 62.5 -> 68.1 | Improved |
| distill final embedding (DisCo) instead of SEED queue similarity | MobileNet-v3-Large student, ResNet-50 teacher · ImageNet linear eval | top-1 accuracy: 55.2 -> 64.4 | Improved |
| combine AT with DisCo | DisCo · ImageNet linear eval | top-1 accuracy: EfficientNet-B0 66.5 -> 66.7, ResNet-18 60.6 -> 60.0 | No clear effect |
| combine KD with DisCo | DisCo · ImageNet linear eval | top-1 accuracy: EfficientNet-B0 66.5 -> 65.8, MobileNet-v3 64.4 -> 65.2 | Mixed |
| use SwAV R50*2 teacher with 200 epochs vs SEED 800 epochs | DisCo EfficientNet-B1 student · ImageNet linear eval | top-1 accuracy: DisCo 64.0 vs SEED 68.0; 290 epochs gives 70.4 | Hurt |
| apply DisCo to DINO ViT-tiny with ViT-small teacher | DINO ViT-tiny · ImageNet linear eval | top-1 accuracy: 63.2 -> 68.4 | Improved |
| apply DisCo to DINO XCiT-tiny with XCiT-small teacher | DINO XCiT-tiny · ImageNet linear eval | top-1 accuracy: 67.0 -> 71.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SEED — SEED: Self-supervised Distillation For Visual Representation (ICLR 2021) · better on 8 of 8 shared results
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
- On the Efficacy of Small Self-Supervised Contrastive Models without Distillation Signals (AAAI 2022 · similarity 0.90)
- Slimmed Asymmetrical Contrastive Learning and Cross Distillation for Lightweight Model Training (NeurIPS 2023 · similarity 0.89)
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.89)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.88)
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.87)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
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.
