Research map / Learning theory and algorithms / Prediction and decision losses
Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity
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 |
|---|---|---|---|
| jointly select salient regions across inputs while encouraging diversity among outputs | image classification with Co-Mixup · CIFAR-100 (PreActResNet18) | Top-1 error rate: 20.62 -> 19.87 | Helped |
| use Co-Mixup rather than the strongest mixup baseline | CIFAR-100 (WRN16-8) · CIFAR-100 | Top-1 error rate: 19.24 -> 19.15 | Helped |
| use Co-Mixup rather than the strongest mixup baseline | CIFAR-100 (ResNeXt29-4-24) · CIFAR-100 | Top-1 error rate: 21.12 -> 19.78 | Helped |
| use Co-Mixup rather than the strongest mixup baseline | Tiny-ImageNet (PreActResNet18) · Tiny-ImageNet | Top-1 error rate: 36.52 -> 35.85 | Helped |
| use Co-Mixup rather than the strongest mixup baseline | ImageNet (ResNet-50) · ImageNet | Top-1 error rate: 22.49 -> 22.39 | Helped |
| use Co-Mixup rather than the strongest mixup baseline | Google commands (VGG-11) · Google commands | Top-1 error rate: ~3.67 -> ~3.54 | Helped |
| use Co-Mixup rather than the strongest baseline | weakly supervised localization · ImageNet | Localization accuracy (%): 55.22 -> 55.32 | Helped |
| use Co-Mixup rather than Vanilla | CIFAR-100 classifiers · CIFAR-100 | ECE (%): 3.9 -> 1.9 | Helped |
| train with Co-Mixup rather than the strongest baseline | ImageNet classifiers evaluated with background replacement · ImageNet background random replacement | Top-1 error rate: 39.23 -> 38.77 | Helped |
| train with Co-Mixup rather than the strongest baseline | ImageNet classifiers evaluated with background noise · ImageNet background Gaussian noise | Top-1 error rate: 26.11 -> 25.89 | Helped |
| mix 3 or 4 inputs instead of 2 in conventional Input, Manifold, or CutMix (ablation) | mixup baselines with PreActResNet18 · CIFAR-100 | Top-1 error rate: 22.43 -> 23.03; 22.43 -> 23.12; 21.64 -> 22.13; 21.64 -> 22.07; 21.29 -> 22.01; 21.29 -> 22.20 | Hurt |
| vary smoothness coefficient beta from 0 to 0.64 (ablation) | Co-Mixup with PreActResNet18 · CIFAR-100 | Top-1 error rate: 20.29; 20.18; 19.87; 20.35; 21.24 | Mixed |
Closest papers by meaning
- Puzzle Mix: Exploiting Saliency and Local Statistics for Optimal Mixup (ICML 2020 · similarity 0.88)
- GuidedMixup: An Efficient Mixup Strategy Guided by Saliency Maps (AAAI 2023 · similarity 0.87)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · similarity 0.85)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 2024 · similarity 0.84)
- Adversarial AutoMixup (ICLR 2024 · similarity 0.84)
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.83)
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.