Research map / Adversarial robustness and security / Certified and empirical robustness
Wasserstein Introspective Neural Networks
Kwonjoon Lee, Weijian Xu, Fan Fan, Zhuowen Tu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 INN's classifier loss with a Wasserstein loss and gradient penalty | Introspective neural networks · CIFAR-10 | Inception score: 1.95 ± .01 -> 4.62 ± .05 | Improved |
| Use a single WINN classifier rather than INNg's 20-CNN configuration | Unsupervised texture modeling · Texture images | visual synthesis quality: 20 | Improved |
| Use WINN-single rather than INNg | Unsupervised face modeling · CelebA | visual synthesis quality: 12 CNNs | Improved |
| Use 5 cascaded classifiers instead of a single classifier (ablation) | WINN · CIFAR-10 | Inception score: 4.62 ± .05 -> 5.58 ± .05 | Improved |
| Add WINN training to a vanilla CNN | Vanilla CNN (4 layers) · MNIST test | error: 0.89 % -> 0.67% | Improved |
| Add WINN training to ResNet-32 | ResNet-32 · MNIST test | error: 0.45% -> 0.48% | Hurt |
| Add WINN training to ResNet-32 | ResNet-32 · SVHN test | error: 4.64% -> 4.50% | Improved |
| Train with WINN rather than standard cross-entropy | Vanilla CNN (4 layers) · MNIST adversarial examples | adversarial error: 32.41% -> 7.99% | Improved |
| Train with WINN rather than standard cross-entropy | ResNet-32 · MNIST adversarial examples | adversarial error: 11.28% -> 2.05% | Improved |
| Train with WINN rather than standard cross-entropy | ResNet-32 · SVHN adversarial examples | adversarial error: 29.29% -> 19.53% | Improved |
| Replace a standard GAN discriminator with an introspective WINN classifier | ResNet discriminator trained on CelebA · CelebA | visual synthesis quality | Improved |
Closest papers by meaning
- Some Theoretical Insights into Wasserstein GANs (JMLR 2021 · similarity 0.88)
- Hidden Convexity of Wasserstein GANs: Interpretable Generative Models with Closed-Form Solutions (ICLR 2022 · similarity 0.88)
- Wasserstein Generative Adversarial Networks (ICML 2017 · similarity 0.88)
- Improving the Improved Training of Wasserstein GANs: A Consistency Term and Its Dual Effect (ICLR 2018 · similarity 0.87)
- Don't let your Discriminator be fooled (ICLR 2019 · similarity 0.87)
- Generative Modeling Using the Sliced Wasserstein Distance (CVPR 2018 · 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.
