Research map / Label-efficient and robust learning / Recognition and long-tail
Chroma-VAE: Mitigating Shortcut Learning with Generative Classifiers
Wanqian Yang, Polina Kirichenko, Micah Goldblum, Andrew Gordon Wilson
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 Chroma-VAE instead of JTT | CelebA Blond/Gender · CelebA Blond/Gender | worst-group accuracy: 28.3 -> 54.4 | Improved |
| Use Chroma-VAE instead of JTT | CelebA Attractive/Smiling · CelebA Attractive/Smiling | worst-group accuracy: 30.0 -> 53.1 | Improved |
| Use Chroma-VAE instead of JTT | MF-Dominoes · MF-Dominoes | worst-group accuracy: 0.81 -> 73.8 | Improved |
| Use Chroma-VAE instead of JTT | Chest X-Ray · Chest X-Ray | worst-group accuracy: 52.3 -> 57.8 | Improved |
| Use Chroma-VAE instead of Naive-Class | Chest X-Ray · Chest X-Ray | average accuracy: 85.0 -> 59.9 | Hurt |
| Train the prediction classifier on z2 rather than use the z1 classifier (ablation) | Chroma-VAE · CelebA synthetic patch D_anti | accuracy: 6.03 -> 87.24 | Improved |
| Add a generative component without separating latent subspaces | Naive-VAE-Class versus Naive-Class · ColoredMNIST D_out | accuracy: 4.45 -> 4.69 | No clear effect |
| Increase beta above 1 in the ELBO (ablation) | Chroma-VAE · ColoredMNIST D_out | accuracy: 72.4 -> 70.3; 72.4 -> 62.8; 72.4 -> 49.4 | Hurt |
| Set beta to 0.5 instead of 1 (ablation) | Chroma-VAE · ColoredMNIST D_out | accuracy: 72.4 -> 71.2 | No clear effect |
| Increase latent dimension (ablation) | Chroma-VAE · ColoredMNIST | reconstruction fidelity | Improved |
| Increase partition fraction z_p (ablation) | Chroma-VAE · ColoredMNIST | z2-classifier predictive performance | Hurt |
| Make the shortcut completely correlated with the label during training | Chroma-VAE · CelebA and MF-Dominoes | worst-group accuracy | Improved |
Closest papers by meaning
- On the Foundations of Shortcut Learning (ICLR 2024 · similarity 0.87)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.85)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.84)
- Navigate Beyond Shortcuts: Debiased Learning Through the Lens of Neural Collapse (CVPR 2024 · similarity 0.84)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.84)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.83)
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.
