Research map / Label-efficient and robust learning / Semi-supervised learning
ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching
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 |
|---|---|---|---|
| Add conditional-entropy cycle-consistency regularization to ALI | ALI · Toy 5-component GMM | MSE: 4.856\pm2.920 -> 0.022\pm0.029 | Helped |
| Add conditional-entropy cycle-consistency regularization to ALI | ALI · Toy 5-component GMM | ICP: 2.776\pm1.516 -> 4.595\pm0.604 | Helped |
| Add cycle-consistency regularization | ALI · MNIST | ICP: 8.749 \pm 0.09 -> 9.279 \pm 0.07 | Helped |
| Add cycle-consistency regularization | ALI · MNIST | MSE: 0.4803 \pm 0.100 -> 0.0803 \pm 0.007 | Helped |
| Add cycle-consistency regularization | ALI · CIFAR-10 | ICP: 5.93 \pm 0.0437 -> 6.015 \pm 0.0284 | Helped |
| Add cycle-consistency regularization | ALI · CIFAR-10 | MSE: 0.672 \pm 0.1129 -> 0.4155 \pm 0.2015 | Helped |
| Use deterministic mappings instead of two stochastic mappings (ablation) | ALI · Toy 5-component GMM | MSE: 4.856\pm2.920 -> 2.399\pm3.605 | Mixed |
| Replace one stochastic mapping with a deterministic mapping (ablation) | ALI · Toy 5-component GMM | MSE: 4.856\pm2.920 -> 3.888\pm7.343 | Mixed |
| Use adversarially learned paired reconstruction instead of an $\ell_2$ reconstruction loss (ablation) | DiscoGAN · Car-to-Car | reconstruction fidelity | Helped |
| Use marginal adversarial reconstruction rather than paired joint adversarial reconstruction (ablation) | DiscoGAN · Car-to-Car | reconstruction fidelity | Hurt |
| Add 1% paired supervision | ALICE · Car-to-Car | angle prediction accuracy: 1% | Helped |
| Increase paired supervision to 10% on only 2 or 6 of 11 angles | ALICE · Car-to-Car | angle prediction accuracy: 10% | Helped |
Closest papers by meaning
- JointGAN: Multi-Domain Joint Distribution Learning with Generative Adversarial Nets (ICML 2018 · similarity 0.89)
- Multivariate-Information Adversarial Ensemble for Scalable Joint Distribution Matching (ICML 2019 · similarity 0.87)
- Cooperative Distribution Alignment via JSD Upper Bound (NeurIPS 2022 · similarity 0.85)
- Adversarial Support Alignment (ICLR 2022 · similarity 0.85)
- Triangle Generative Adversarial Networks (NeurIPS 2017 · similarity 0.85)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.84)
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.