Research map / Label-efficient and robust learning / OOD and anomaly detection
Embedding Contrastive Unsupervised Features to Cluster in- and Out-of-Distribution Noise in Corrupted Image Datasets
Paul Albert, Eric Arazo, Noel E. O’Connor, Kevin McGuinness
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 mixup (ablation) | SNCF · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 61.66 -> 67.62 | Improved |
| add spectral embedding of unsupervised features before clustering (ID corr and rm OOD, no contrastive) (ablation) | SNCF · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 54.81 -> 61.83 | Improved |
| add guided contrastive objective on top of mixup without noise correction (ablation) | cross-entropy with mixup · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 59.54 -> 62.83 | Improved |
| use OOD samples in guided contrastive loss instead of removing them (ablation) | SNCF with embedding · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 64.51 -> 65.22 | Improved |
| add equal sampling of clean, ID noisy and OOD samples per mini-batch (ablation) | SNCF · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 65.22 -> 67.62 | Improved |
| add guided contrastive loss with embedding and ID correction (ablation) | SNCF · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 61.40 -> 63.91 | Improved |
| use ID correction only without embedding or contrastive loss (ablation) | cross-entropy · CIFAR-100 + ImageNet32 OOD (r_out 0.4, r_in 0.2) | best accuracy: 58.94 -> 57.02 | Hurt |
| raise xi in OPTICS above 0.03 (ablation) | SNCF clustering · CIFAR-100 + ImageNet32 | clustering accuracy | Hurt |
| replace Places365 with ImageNet32 as OOD corruption | SNCF · CIFAR-100 (r_out 0.4, r_in 0.4) | best accuracy: 54.04 vs 50.95 | Mixed |
| use SNCF instead of RRL at high noise | RRL · CIFAR-100 + ImageNet32 OOD (r_out 0.6, r_in 0.2) | best accuracy: 26.76 -> 53.26 | Improved |
| use SNCF instead of RRL on Places365 at low noise | RRL · CIFAR-100 + Places365 OOD (r_out 0.2, r_in 0.2) | best accuracy: 72.62 -> 71.25 | Hurt |
| use SNCF instead of PropMix on CNWL | PropMix · CNWL miniImageNet 32x32 (noise 40%) | best accuracy: 56.22 -> 59.94 | Improved |
Closest papers by meaning
- An accurate detection is not all you need to combat label noise in web-noisy datasets (ECCV 2024 · similarity 0.90)
- Learning From Noisy Data With Robust Representation Learning (ICCV 2021 · similarity 0.88)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.87)
- Provable Discriminative Hyperspherical Embedding for Out-of-Distribution Detection (AAAI 2025 · similarity 0.86)
- Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport (AAAI 2024 · similarity 0.85)
- Enhanced OoD Detection through Cross-Modal Alignment of Multi-Modal Representations (CVPR 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.
