Research map / Label-efficient and robust learning / Semi-supervised learning
Harnessing Hard Mixed Samples with Decoupled Regularizer
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 |
|---|---|---|---|
| replace MCE with DM(CE) loss for static mixup | Mixup/CutMix/ManifoldMix/FMix/ResizeMix · CIFAR-100 and Tiny-ImageNet | top-1 accuracy: +0.78, +0.77, +0.34 (CIFAR-100), +1.18, +1.62 (Tiny) avg | Helped |
| replace MCE with DM(CE) for ResizeMix | ResizeMix · CIFAR-100 WRN-28-8 | top-1 accuracy: 84.87 -> 84.72 | Hurt |
| add DM(CE) to dynamic mixup (PuzzleMix, AutoMix, SAMix) | dynamic mixup · CIFAR-100 / Tiny-ImageNet | top-1 accuracy: +0.15~0.4 (CIFAR-100), +0.56~0.95 (Tiny) avg | Helped |
| add DM(CE) and AS to Self-Tuning with Mixup | Self-Tuning · CUB-200 15% labels | top-1 accuracy: 62.38 -> 73.06 | Helped |
| add DM(CE)+AS to FixMatch+Mixup | FixMatch · CIFAR-100 400 labels | top-1 accuracy: 50.61 -> 59.75 | Helped |
| apply plain Mixup (MCE) on top of FixMatch | FixMatch · CIFAR-100 400 labels | top-1 accuracy: 53.58 -> 50.61 | Hurt |
| add DM to MixMatch | MixMatch · CIFAR-10 250 labels | top-1 accuracy: 86.37 -> 89.16 | Helped |
| asymmetric strategy AS (lambda<=0.5) instead of MCE or AS(lambda>=0.5) (ablation) | Self-Tuning SSL transfer · CUB-200 R-18 | top-1 accuracy | Helped |
| add DM(CE) on top of AS (ablation) | Self-Tuning with AS · CUB-200 R-18 | top-1 accuracy | Helped |
| set eta to 1 for dynamic mixup vs 0.1 for static mixup (ablation) | DM(CE) · ImageNet-1k ResNet-34 | top-1 accuracy | Mixed |
| tune DM(BCE) xi,t per mixup type (cut: xi=0.8,t=1; interpolation: xi=1.0,t=0.5) (ablation) | DM(BCE) · ImageNet-1k ResNet-34 | top-1 accuracy | Mixed |
| use two-hot encoding (MBCE(two)) instead of DM(BCE) rescaling (ablation) | MBCE · ImageNet-1k RSB A3 | top-1 accuracy | Hurt |
Closest papers by meaning
- Over-Training with Mixup May Hurt Generalization (ICLR 2023 · similarity 0.90)
- C-Mixup: Improving Generalization in Regression (NeurIPS 2022 · similarity 0.90)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · similarity 0.88)
- Adversarial AutoMixup (ICLR 2024 · similarity 0.88)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 2024 · similarity 0.88)
- Selective Mixup Helps with Distribution Shifts, But Not (Only) because of Mixup (ICML 2024 · 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.