Research map / Label-efficient and robust learning / Recognition and long-tail
On Predicting Generalization using GANs
Yi Zhang, Arushi Gupta, Nikunj Saunshi, Sanjeev Arora
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 |
|---|---|---|---|
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 1: VGG on CIFAR-10 | Conditional Mutual Information: 52.59 -> 62.62 | Improved |
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 2: NIN on SVHN | Conditional Mutual Information: 48.34 -> 34.72 | Hurt |
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 4: AllConv on CINIC-10 | Conditional Mutual Information: 57.81 -> 52.80 | Hurt |
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 5: AllConv on CINIC-10 | Conditional Mutual Information: 34.57 -> 53.56 | Improved |
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 8: VGG on F-MNIST | Conditional Mutual Information: 16.23 -> 30.25 | Improved |
| Use GAN-generated labeled samples to predict classifier generalization | Deep classifiers on PGDL tasks · PGDL Task 9: NIN on CIFAR-10 | Conditional Mutual Information: 25.86 -> 33.51 | Improved |
| Use BigGAN+DiffAug synthetic examples rather than the qrt+log predictor | DEMOGEN generalization prediction · NiN+CIFAR-10 | adjusted R²: 0.94 -> 0.995 | Improved |
| Use BigGAN+DiffAug synthetic examples rather than the qrt+log predictor | DEMOGEN generalization prediction · ResNet+CIFAR-10 | adjusted R²: 0.87 -> 0.981 | Improved |
| Use BigGAN+DiffAug synthetic examples rather than the qrt+log predictor | DEMOGEN generalization prediction · ResNet+CIFAR-100 | adjusted R²: 0.97 -> 0.993 | Improved |
| Apply ordinary data augmentation to real GAN-training examples only (ablation) | BigGAN generalization predictor · DEMOGEN CIFAR-100 | test accuracy versus synthetic accuracy agreement | Hurt |
| Remove data augmentation from GAN training (ablation) | BigGAN generalization predictor · DEMOGEN CIFAR-100 | test accuracy versus synthetic accuracy agreement | Improved |
| Apply differentiable augmentation to both real and fake GAN images (ablation) | BigGAN generalization predictor · DEMOGEN CIFAR-100 | test accuracy versus synthetic accuracy agreement | Improved |
Closest papers by meaning
- On the Discrimination-Generalization Tradeoff in GANs (ICLR 2018 · similarity 0.89)
- Improving Generalization and Stability of Generative Adversarial Networks (ICLR 2019 · similarity 0.89)
- On the Generalization Analysis of Adversarial Learning (ICML 2022 · similarity 0.88)
- A Large-Scale Study on Regularization and Normalization in GANs (ICML 2019 · similarity 0.87)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.85)
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · 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.
