Research map / Label-efficient and robust learning / Recognition and long-tail
Robust Contrastive Learning Using Negative Samples with Diminished Semantics
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 |
|---|---|---|---|
| Add texture-based non-semantic negatives | MoCo-v2, k=16384 · ImageNet-100 | top-1 accuracy: 77.88 -> 77.76 | No clear effect |
| Add texture-based non-semantic negatives | MoCo-v2, k=16384 · ImageNet-C | top-1 accuracy: 43.08 -> 43.58 | Helped |
| Add patch-based non-semantic negatives with α=2 | MoCo-v2, k=16384 · ImageNet-100 | top-1 accuracy: 77.88 -> 79.35 | Helped |
| Add patch-based non-semantic negatives with α=2 | MoCo-v2, k=16384 · ImageNet-C | top-1 accuracy: 43.08 -> 45.13 | Helped |
| Add patch-based non-semantic negatives with α=3 | MoCo-v2, k=16384 · ImageNet-S | top-1 accuracy: 28.24 -> 34.03; +5.79% | Helped |
| Add patch-based non-semantic negatives with α=3 | MoCo-v2, k=16384 · ImageNet-R | top-1 accuracy: 32.92 -> 36.89; +5.97% | Helped |
| Add patch-based non-semantic negatives | MoCo-v2, k=8192 · ImageNet-100 | top-1 accuracy: 77.73 -> 79.54 | Helped |
| Add patch-based non-semantic negatives | BYOL · ImageNet-R | top-1 accuracy: 39.53 -> 41.16 | Helped |
| Add patch-based non-semantic negatives | MoCo-v2 · ImageNet-S | top-1 accuracy: 17.47 -> 18.58 | Helped |
| Add patch-based non-semantic negatives | MoCo-v2 · ImageNet-C | mCE: 87.7 -> 87.6 | Helped |
| Add patch-based non-semantic negatives | MoCo-v2 · STL-10 | top-1 accuracy: 88.00 -> 89.36 | Helped |
| Use α=3 rather than α=2 for patch-based negatives (ablation) | MoCo-v2, k=16384 · ImageNet-100 | top-1 accuracy: 79.35 -> 75.58 | Hurt |
Closest papers by meaning
- Conditional Negative Sampling for Contrastive Learning of Visual Representations (ICLR 2021 · similarity 0.91)
- Contrastive Learning with Hard Negative Samples (ICLR 2021 · similarity 0.90)
- Hard Negative Mixing for Contrastive Learning (NeurIPS 2020 · similarity 0.89)
- Diversity Is All You Need for Contrastive Learning: Spectral Bounds on Gradient Magnitudes (NeurIPS 2025 · similarity 0.88)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 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.