Research map / Label-efficient and robust learning / Self-supervised and contrastive
How Well Does Self-Supervised Pre-Training Perform with Streaming Data?
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 |
|---|---|---|---|
| Use sequential rather than joint self-supervised pre-training on IID or moderately shifted chunks | MoCo-v2 pre-training · ImageNet instance and random class incremental sequences | linear accuracy gap (%): 0.41, 1.02, 1.04; 0.42, 0.94, 1.13 | No clear effect |
| Use sequential self-supervised pre-training on distant class chunks | MoCo-v2 pre-training · ImageNet distant class incremental sequence | linear accuracy gap (%): 2.34, 3.81, 4.62 | Hurt |
| Use sequential rather than joint supervised pre-training | supervised ResNet50 pre-training · ImageNet instance, random class, and distant class incremental sequences | linear accuracy gap (%): 2.26, 3.27, 4.83; 5.63, 8.73, 10.68; 7.77, 12.50, 15.75 | Hurt |
| Add MAS regularization to sequential SSL (ablation) | MoCo-v2 SSL-ST · ImageNet distant class incremental sequence | linear accuracy gap (%): 2.34 -> 1.82; 3.81 -> 2.73; 4.62 -> 3.17 | Helped |
| Add MAS regularization and data replay to sequential SSL (ablation) | MoCo-v2 SSL-ST · ImageNet distant class incremental sequence | linear accuracy gap (%): 2.34 -> 1.47; 3.81 -> 2.01; 4.62 -> 2.10 | Helped |
| Add 10% data replay to sequential SSL (ablation) | MoCo-v2 SSL-ST · Domain incremental sequence | downstream transfer performance gap to joint SSL | Helped |
| Use sequential SSL instead of sequential supervised learning | ImageNet streaming pre-training · Instance, random class, and distant class incremental sequences | BWT Top-1 (%): -9.45 -> 3.61; -20.63 -> -5.17; -40.43 -> -13.24 | Helped |
| Use sequential SSL rather than joint SSL on severe distribution shifts | MoCo-v2 pre-training · Distant class and domain incremental sequences | downstream transfer performance | Hurt |
| Use sequential SSL rather than joint SSL on negligible or moderate distribution shifts | MoCo-v2 pre-training · Instance and random class incremental sequences | downstream transfer performance | No clear effect |
| Use BYOL sequentially rather than jointly | BYOL pre-training · ImageNet distant class incremental sequence | downstream transfer performance | Hurt |
| Increase the sequence length | Sequential SSL pre-training · 8-chunk ImageNet-2K sequence | performance gap to joint training | Hurt |
| Add 10% previous-chunk data replay (ablation) | Sequential SSL pre-training · 8-chunk ImageNet-2K sequence | performance gap to joint training | Helped |
Closest papers by meaning
- Rethinking Supervised Pre-Training for Better Downstream Transferring (ICLR 2022 · similarity 0.89)
- Learning from Streaming Video with Orthogonal Gradients (CVPR 2025 · similarity 0.86)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · similarity 0.86)
- Memorization in Self-Supervised Learning Improves Downstream Generalization (ICLR 2024 · similarity 0.86)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.86)
- Beyond Random Augmentations: Pretraining with Hard Views (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.