Research map / Adversarial robustness and security / Certified and empirical robustness
PatchUp: A Feature-Space Block-Level Regularization Technique for Convolutional Neural Networks
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 |
|---|---|---|---|
| Use feature-space block mixing or swapping with PatchUp | CNN classifiers · CIFAR-10, CIFAR-100, SVHN, Tiny-ImageNet, and ImageNet2012 | classification error: 2.92 ± 0.13 vs 3.16 ± 0.11 on CIFAR-10; 19.12 ± 0.17 vs 20.65 ± 0.21 on CIFAR-100 | Helped |
| Use Soft rather than Hard PatchUp | CNN classifiers · Tiny-ImageNet | error: 38.68 ± 0.34 vs 40.55 ± 0.15 on ResNet101 | Helped |
| Use Soft rather than Hard PatchUp | PreActResNet34 on CIFAR100 · affine-transformed CIFAR100 | error rate: 31.282 ± 0.622 vs 31.340 ± 0.318 for Rotate (-20, 20); 52.648 ± 0.616 vs 46.924 ± 1.035 for Scale (0.6) | Mixed |
| Add L2 alongside L1 in the Soft PatchUp objective (ablation) | Soft PatchUp · CIFAR-100 validation | error rate: 16.86 ± 0.67 -> 16.75 ± 0.29 | Helped |
| Add L2 alongside L1 in the Hard PatchUp objective (ablation) | Hard PatchUp · CIFAR-100 validation | error rate: 16.14 ± 0.23 -> 16.02 ± 0.36 | Helped |
| Select the PatchUp layer randomly rather than fixing it to one layer (ablation) | Hard PatchUp on WRN-28-10 · CIFAR-100 | test error: 17.86 ± 0.16 -> 16.13 ± 0.20; 21.42 ± 0.28 -> 16.13 ± 0.20; 25.25 ± 0.14 -> 16.13 ± 0.20 | Helped |
| Use PatchUp masks in input space (ablation) | Hard PatchUp on WRN-28-10 · CIFAR-100 | test error: 16.13 ± 0.20 -> 16.90 ± 0.0.21 | Hurt |
Closest papers by meaning
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 2024 · similarity 0.88)
- Kernel-convoluted Deep Neural Networks with Data Augmentation (AAAI 2021 · similarity 0.88)
- Efficient Learning of CNNs using Patch Based Features (ICML 2022 · similarity 0.88)
- C-Mixup: Improving Generalization in Regression (NeurIPS 2022 · similarity 0.87)
- Harnessing Hard Mixed Samples with Decoupled Regularizer (NeurIPS 2023 · similarity 0.87)
- Pushing Boundaries: Mixup's Influence on Neural Collapse (ICLR 2024 · 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.