Research map / Label-efficient and robust learning / Domain adaptation and generalization
Normalized Wasserstein for Mixture Distributions With Applications in Adversarial Learning and Domain Adaptation
Yogesh Balaji, Rama Chellappa, Soheil Feizi
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 Wasserstein distance with normalized Wasserstein (NW) for domain adaptation | domain adaptation · MNIST->MNIST-M (3 modes, imbalanced) | accuracy (%): 61.75 -> 75.06 | Improved |
| replace Wasserstein distance with NW for domain adaptation | domain adaptation · MNIST->MNIST-M (5 modes, imbalanced) | accuracy (%): 60.56 -> 76.16 | Improved |
| replace Wasserstein distance with NW for domain adaptation | domain adaptation · MNIST->MNIST-M (10 modes, imbalanced) | accuracy (%): 58.22 -> 68.57 | Improved |
| apply DANN or Wasserstein alignment instead of no adaptation under imbalanced modes | source-only baseline · MNIST->MNIST-M (3 modes, imbalanced) | accuracy (%): 66.63 -> 62.34 | Hurt |
| replace Wasserstein distance with NW on pretrained ResNet-18 | domain adaptation · VISDA synthetic-to-real (3 classes, imbalanced) | accuracy (%): 64.84 -> 73.23 | Improved |
| use NW instead of Wasserstein when modes are balanced | domain adaptation · MNIST->MNIST-M (mode-balanced, full data) | accuracy (%): 83.47 -> 84.16 | No clear effect |
| use NW instead of Wasserstein when modes are balanced | domain adaptation · VISDA (3 classes, mode-balanced) | accuracy (%): 90.08 -> 90.72 | No clear effect |
| replace Wasserstein distance with NW for denoising adaptation | autoencoder domain adaptation · MNIST denoising (target reconstruction) | target L2 reconstruction error: 0.52 -> 0.18 | Improved |
| learn mixture proportions (NWGAN) instead of fixed uniform proportions (MGAN) (ablation) | mixture GAN · Mixture of 9 Gaussians | pi estimation error: 0.7157 -> 0.0001 | Improved |
| optimize pi instead of fixing pi to 1/k (ablation) | NWGAN k=2 · CIFAR-10 + CelebA mixture | disentanglement of modes | Improved |
| use NWGAN with 4 modes instead of single WGAN-GP generator | WGAN-GP · CIFAR-10 | inception score: 7.56 -> 7.89 | Improved |
| replace k-means with NW adversarial clustering | clustering · Imbalanced MNIST (digits 2,4,6) clustering | cluster purity: 0.82 -> 0.98 | Improved |
Closest papers by meaning
- Understanding Entropic Regularization in GANs (JMLR 2024 · similarity 0.88)
- Generative Modeling Using the Sliced Wasserstein Distance (CVPR 2018 · similarity 0.88)
- Wasserstein of Wasserstein Loss for Learning Generative Models (ICML 2019 · similarity 0.87)
- Max-Sliced Wasserstein Distance and Its Use for GANs (CVPR 2019 · similarity 0.87)
- Approximability of Discriminators Implies Diversity in GANs (ICLR 2019 · similarity 0.86)
- A Two-Step Computation of the Exact GAN Wasserstein Distance (ICML 2018 · similarity 0.86)
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.
