Research map / Label-efficient and robust learning / Self-supervised and contrastive
Hard Negative Mixing for Contrastive Learning
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 hard-negative mixing to MoCo-v2 | MoCo-v2 · ImageNet-100, 200 epochs | Top-1 accuracy: 78.0 (± 0.2) -> 79.0 (± 0.4) | Helped |
| Add hard-negative mixing to MoCo-v2 | MoCo-v2 · PASCAL VOC 2007, 100-epoch pretraining | AP: 53.7 -> 54.7 | Helped |
| Add hard-negative mixing to MoCo-v2 | MoCo-v2 · PASCAL VOC 2007, 200-epoch pretraining | AP: 56.8 -> 57.5 | Helped |
| Add hard-negative mixing to MoCo-v2 | MoCo-v2 · ImageNet-1K, 200 epochs | Top-1 accuracy: 67.9 -> 67.6 | No clear effect |
| Add hard-negative mixing to MoCo-v2 | MoCo-v2 · COCO val2017, 100-epoch pretraining | AP (bounding box): 37.0 -> 37.8 | Helped |
| Add hard-negative mixing to MoCo-v2 | MoCo-v2 · COCO val2017, 100-epoch pretraining | AP (instance mask): 32.7 -> 33.2 | Helped |
| Use pairwise hard-negative mixing only (ablation) | MoCo-v2 · ImageNet-100, 200 epochs | Top-1 accuracy: 78.0 -> 78.8 | Helped |
| Use query-negative mixing only (ablation) | MoCo-v2 · ImageNet-100, 200 epochs | Top-1 accuracy: 78.0 -> 79.0 | Helped |
| Select hard negatives using the query rather than the key (ablation) | MoCHi · ImageNet-100 | Top-1 accuracy: about 0.2% lower with key-based ranking | Helped |
| Mix the query with hard negatives instead of mixing the key with hard negatives (ablation) | MoCHi · ImageNet-100 | Top-1 accuracy: about 0.1-0.2% lower with key mixing | Helped |
| Weight synthetic negative logits by the query mixing coefficient (ablation) | MoCHi · ImageNet-100 | Top-1 accuracy | Hurt |
| Sample negatives non-uniformly using a softmax over the top N (ablation) | MoCHi · ImageNet-100 | Top-1 accuracy: 10-30% | Hurt |
Closest papers by meaning
- Conditional Negative Sampling for Contrastive Learning of Visual Representations (ICLR 2021 · similarity 0.91)
- Incremental False Negative Detection for Contrastive Learning (ICLR 2022 · similarity 0.90)
- Contrastive Learning with Hard Negative Samples (ICLR 2021 · similarity 0.89)
- Robust Contrastive Learning Using Negative Samples with Diminished Semantics (NeurIPS 2021 · similarity 0.89)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · 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.