Research map / Adversarial robustness and security / Certified and empirical robustness
Manifold Mixup: Better Representations by Interpolating Hidden States
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 |
|---|---|---|---|
| interpolate hidden representations and their soft labels at randomly selected layers | PreActResNet18 · CIFAR-10 | test error (%): 3.82±0.048 -> 2.95 ± 0.046 | Helped |
| interpolate hidden representations and their soft labels at randomly selected layers | PreActResNet18 · CIFAR-10 | Test NLL: 0.186 ± 0.004 -> 0.137 ± 0.003 | Helped |
| interpolate hidden representations rather than only inputs | PreActResNet18 · CIFAR-100 | test error (%): 22.11±0.424 -> 20.34±0.525 | Helped |
| interpolate hidden representations rather than only inputs | PreActResNet18 · SVHN | test error (%): 2.76±0.014 -> 2.27±0.011 | Helped |
| use hidden-state interpolation instead of input Mixup | PreActResNet18 · CIFAR-10 FGSM | test accuracy: 71.51 -> 77.50 | Helped |
| block gradients before the layer where mixing is applied (ablation) | Manifold Mixup on PreActResNet18 · CIFAR-10 | test error (%): 4.33% | Hurt |
| restrict eligible mixing layers to the input layer {0} instead of {0,1,2} (ablation) | Manifold Mixup · CIFAR-10 | test accuracy: 97.23 -> 96.73 | Hurt |
| remove all eligible mixing layers (ablation) | Manifold Mixup · CIFAR-10 | test accuracy: 97.23 -> 95.27 | Hurt |
| increase α from 0.5 to 2.0 (ablation) | Manifold Mixup · CIFAR-10 | test accuracy: 96.76 -> 97.23 | Helped |
| mix in hidden states instead of input pixels | PreActResNet34 · CIFAR-100 test split with Zoom In (60% rescale) | test accuracy: 13.75 -> 11.49 | Hurt |
| use Manifold Mixup rather than Input Mixup | fully-connected MNIST network · MNIST first hidden layer representations | sum of non-largest singular values: 66.89 -> 40.98 | Helped |
| train with Manifold Mixup rather than PGD training | PreActResNet18 · CIFAR-10 FGSM | test accuracy: 56.10 -> 77.50 | Helped |
Closest papers by meaning
- MixUp as Locally Linear Out-of-Manifold Regularization (AAAI 2019 · similarity 0.90)
- Nonlinear Mixup: Out-Of-Manifold Data Augmentation for Text Classification (AAAI 2020 · similarity 0.87)
- PatchUp: A Feature-Space Block-Level Regularization Technique for Convolutional Neural Networks (AAAI 2022 · similarity 0.87)
- Approximating Latent Manifolds in Neural Networks via Vanishing Ideals (ICML 2025 · similarity 0.86)
- Hardness of Learning Neural Networks under the Manifold Hypothesis (NeurIPS 2024 · similarity 0.85)
- Adversarial Purification with the Manifold Hypothesis (AAAI 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.