Research map / Label-efficient and robust learning / Domain adaptation and generalization
Controlling Directions Orthogonal to a Classifier
Yilun Xu, Hao He, Tianxiao Shen, Tommi S. Jaakkola
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 |
|---|---|---|---|
| Add classifier orthogonalization to CycleGAN using the Fish principal classifier | CycleGAN · CMNIST | Z2 accuracy: 14.3 -> 98.0 | Improved |
| Add classifier orthogonalization using an oracle principal classifier | CycleGAN · CelebA-GH | Z2 accuracy: 16.8 -> 57.4 | Improved |
| Add classifier orthogonalization using an oracle principal classifier | CycleGAN · CelebA-GH | FID: 38.5 -> 39.7 | Hurt |
| Orthogonalize the domain discriminator with respect to task labels | VADA · Visual domain adaptation with label shift | test accuracy: VADA + classifier orthogonalization improves over VADA across all tasks | Improved |
| Use importance-weighted domain adaptation rather than the original VADA objective | VADA · MNIST → MNIST-M and MNIST → SVHN | test accuracy: 77.8 -> 71.2; 35.7 -> 34.5 | Hurt |
| Orthogonalize the ERM classifier with respect to gender | ERM classifier · UCI Adult | ΔEO: 0.20 -> 0.12 | Improved |
| Orthogonalize the ERM classifier with respect to gender | ERM classifier · UCI Adult | accuracy: 84.5 -> 81.6 | Hurt |
| Orthogonalize the ERM classifier with respect to age | ERM classifier · UCI German credit | ΔDP: 0.19 -> 0.09 | Improved |
| Orthogonalize the ERM classifier with respect to age | ERM classifier · UCI German credit | ΔEO: 0.33 -> 0.18 | Improved |
| Orthogonalize the ERM classifier with respect to age | ERM classifier · UCI German credit | accuracy: 76.0 -> 75.4 | Hurt |
| Train the orthogonal classifier by importance sampling instead of classifier orthogonalization | Orthogonal classifier · Synthetic variation in latent divergence and learning rate | test loss | Hurt |
Closest papers by meaning
- Project and Probe: Sample-Efficient Adaptation by Interpolating Orthogonal Features (ICLR 2024 · similarity 0.82)
- Fixed Non-negative Orthogonal Classifier: Inducing Zero-mean Neural Collapse with Feature Dimension Separation (ICLR 2024 · similarity 0.81)
- Style Neophile: Constantly Seeking Novel Styles for Domain Generalization (CVPR 2022 · similarity 0.81)
- $\boldsymbol{\lambda}$-Orthogonality Regularization for Compatible Representation Learning (NeurIPS 2025 · similarity 0.80)
- Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments (AAAI 2024 · similarity 0.79)
- Exploring Orthogonality in Open World Object Detection (CVPR 2024 · similarity 0.79)
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.
