Research map / Label-efficient and robust learning / Recognition and long-tail
Conditional Negative Sampling for Contrastive Learning of Visual Representations
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 |
|---|---|---|---|
| sample negatives within a similarity ring rather than uniformly | IR · CIFAR10, CIFAR100, STL10, ImageNet | Transfer Acc.: 81.2 -> 83.9; 60.4 -> 62.3; 61.4 -> 64.3; 43.2 -> 48.4 | Helped |
| sample negatives within a similarity ring rather than uniformly | CMC · CIFAR10, CIFAR100, STL10, ImageNet | Transfer Acc.: 85.6 -> 87.6; 56.0 -> 56.0; 63.8 -> 66.4; 48.2 -> 50.4 | Mixed |
| sample negatives within a similarity ring rather than uniformly | MoCo · CIFAR10, CIFAR100, STL10, ImageNet | Transfer Acc.: 83.1 -> 86.1; 59.1 -> 61.5; 63.8 -> 65.2; 52.8 -> 54.6 | Helped |
| remove the annealing schedule (ablation) | IRing · CIFAR10 | Acc.: 83.9 -> 81.4 | Hurt |
| remove the annealing schedule (ablation) | IRing · ImageNet | Acc.: 48.4 -> 41.3 | Hurt |
| include the closest examples by setting w_u to 100 (ablation) | IRing · CIFAR10 | Acc.: 83.9 -> 82.1 | Hurt |
| include the closest examples by setting w_u to 100 (ablation) | IRing · ImageNet | Acc.: 48.4 -> 47.3 | Hurt |
| use Ring negatives for CIFAR10-pretrained features | IR · Meta-Dataset transfer classification | accuracy: Aircraft 40.9 -> 40.6; CUBirds 17.9 -> 17.9; DTD 39.2 -> 39.5; Fungi 2.7 -> 3.4; MNIST 96.9 -> 97.8; FashionMNIST 91.7 - | Mixed |
| use Ring negatives for CIFAR10-pretrained features | MoCo · Meta-Dataset transfer classification | accuracy: Aircraft 41.5 -> 41.6; CUBirds 18.0 -> 18.6; DTD 39.7 -> 39.5; Fungi 3.1 -> 3.6; MNIST 96.9 -> 97.9; FashionMNIST 90.9 - | Mixed |
| use Ring negatives for CIFAR10-pretrained features | CMC · Meta-Dataset transfer classification | accuracy: Aircraft 40.1 -> 40.8; CUBirds 15.8 -> 16.8; DTD 38.3 -> 40.6; Fungi 4.3 -> 4.2; MNIST 97.5 -> 97.9; FashionMNIST 91.5 - | Mixed |
| use Ring negatives | IR · COCO object detection, instance segmentation and keypoint detection; VOC object detection | AP: COCO detection AP 8.6 -> 10.9; COCO segmentation AP 8.5 -> 11.0; COCO keypoint AP 34.6 -> 37.2; VOC detection AP 5.5 -> | Helped |
| use Ring negatives | MoCo · COCO object detection, instance segmentation and keypoint detection; VOC object detection | AP: COCO detection AP 6.0 -> 9.4; COCO segmentation AP 10.8 -> 12.0; COCO keypoint AP 37.6 -> 38.7; VOC detection AP 7.3 -> | Helped |
Closest papers by meaning
- Robust Contrastive Learning Using Negative Samples with Diminished Semantics (NeurIPS 2021 · similarity 0.91)
- Hard Negative Mixing for Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.90)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.88)
- Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 2024 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.88)
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.