Research map / Label-efficient and robust learning / Recognition and long-tail
Adversarial AutoMixup
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AdAutoMixup, an adversarial automatic mixup method where an attention-based generator mixes N images to create hard samples that challenge a classifier, stabilized by an EMA teacher and cosine similarity.
Problem
Automatic mixup methods such as AutoMix optimize mixed-sample generation and classification toward the same objective, so generated mixed samples are simple and not diverse, leading to overfitting on image classification.
Approach
Train the mixup generator adversarially (gradient ascent on classifier loss) against the target classifier so it produces hard samples, mix any N images via cross-attention rather than only two, and use an EMA teacher plus cosine similarity to keep mixed images semantically meaningful.
Key results
- On CIFAR-100, AdAutoMix reaches 82.32% (ResNet18) and 84.22% (ResNeXt50), beating the second-best AutoMix by 0.28 and 0.58 points.
- On Tiny-ImageNet, AdAutoMix improves over the second-best method by 1.86 points (ResNet18) and 2.17 points (ResNeXt50).
- On ImageNet-1K, AdAutoMix gains are small: 70.86% vs 70.50% for ResNet18 and 78.04% vs 77.91% for ResNet50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace AutoMix with adversarial automatic mixup | ResNet18 classifier · CIFAR100 | top-1 accuracy: +0.28 | Helped |
| Replace AutoMix with adversarial automatic mixup | ResNeXt50 classifier · Tiny-ImageNet | top-1 accuracy: +2.17 | Helped |
| Replace AutoMix with adversarial automatic mixup | ResNet50 classifier · CUB-200 | accuracy: +0.69 | Helped |
| Replace AutoMix with adversarial automatic mixup | ResNet18 classifier · CIFAR100-C | accuracy: 0.40% | Helped |
| Replace AutoMix with adversarial automatic mixup | ResNet18 classifier · CIFAR100 | ECE: -0.2 | Helped |
| Add 0.5 L_ace (ablation) | Base AdAutoMix · CIFAR100 | top-1 accuracy: 79.38 -> 79.98 | Helped |
| Add 0.5 L_mce alongside 0.5 L_ace (ablation) | Base AdAutoMix + 0.5 L_ace · CIFAR100 | top-1 accuracy: 79.98 -> 80.04 | Helped |
| Add -0.3 L_amce (ablation) | Base + 0.5 L_ace + 0.5 L_mce · CIFAR100 | top-1 accuracy: 80.04 -> 81.32 | Helped |
| Add 0.7 L_cosine (ablation) | Base + 0.5 L_ace + 0.5 L_mce - 0.3 L_amce · CIFAR100 | top-1 accuracy: 81.32 -> 81.55 | Helped |
| Set mixing ratio λ to 1.0 rather than 0.2, 2.0, 5.0, or 10.0 (ablation) | AdAutoMix ResNet18 · CIFAR100 | accuracy: 82.27; 82.32; 81.73; 80.02; 81.05 | Mixed |
| Set input image count N to 3 rather than N=1, 2, 4, or 5 (ablation) | AdAutoMix ResNet18 · CIFAR100 | Top1-Acc: 78.04; 82.16; 82.32; 81.78; 80.79 | Mixed |
| Use adversarial training (ablation) | AdAutoMix ResNet18 · CIFAR100 | classification accuracy | Helped |
Surprising findings
- Reported ablation shows the adversarial EMA-teacher and cosine terms give most of the gain, with the adversarial training alone changing results only modestly in the appendix comparisons.
- Gains on ImageNet-1K are marginal (0.13 to 0.36 points), and the method takes more training time than prior approaches.
Limitations
- Higher training cost per iteration than prior automatic mixup methods.
- Gains on ImageNet-1K are marginal.
- Hyperparameters (alpha, beta, lambda, N) were tuned on CIFAR-100 with ResNet18, which may limit transfer to other settings.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet ImageNet-1K CUB-200-2011 FGVC-Aircraft Stanford Cars CIFAR-100-C
Benchmarks: CIFAR-100 Tiny-ImageNet ImageNet-1K CUB-200 FGVC-Aircraft Stanford Cars CIFAR-100-C
Metrics: top-1 accuracy expected calibration error FGSM error occlusion accuracy
Baselines: Mixup CutMix ManifoldMix FMix ResizeMix SaliencyMix PuzzleMix AutoMix
Builds on: AutoMix Mixup CutMix OpenMixup
Closest papers by meaning
- A Unified Analysis of Mixed Sample Data Augmentation: A Loss Function Perspective (NeurIPS 2022 · similarity 0.91)
- Adversarial Vertex Mixup: Toward Better Adversarially Robust Generalization (CVPR 2020 · similarity 0.89)
- StyleMix: Separating Content and Style for Enhanced Data Augmentation (CVPR 2021 · similarity 0.89)
- mixup: Beyond Empirical Risk Minimization (ICLR 2018 · similarity 0.88)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 2024 · similarity 0.87)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 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.