Research map / Label-efficient and robust learning / Domain adaptation and generalization
Symmetric Self-Paced Learning for Domain Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Symmetric Self-Paced Learning (SSPL), a curriculum-style training scheduler plus a gradient-magnitude difficulty measure (GDM) that can wrap existing domain generalization methods to improve unseen-domain accuracy.
Problem
Domain generalization methods present training examples in random order and ignore example difficulty; conventional curriculum and self-paced schedules bias training toward easy examples from small-gap domains.
Approach
A symmetric scheduler whose per-example weights start high for easy examples and reverse by the final epoch so hard examples gain emphasis, combined with a difficulty measure based on the gradient norm with respect to the input rather than the training loss.
Key results
- SSPL improves ERM on Digits by up to 16.31% on MNIST-M and 5.84% on SYN with random rotation.
- On PACS, SSPL gains up to 10.41% on Sketch over the ERM baseline.
- On VLCS, SSPL gains up to 7.49% on Sun; classic SPL with ERM drops accuracy substantially on several domains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SSPL to ERM | ERM · Digits (MNIST, MNIST-M, SVHN, SYN) | accuracy: 96.4 ± .2 -> 96.9 ± .1; 62.6 ± .1 -> 65.5 ± .4; 66.7 ± .1 -> 67.7 ± .4; 83.8 ± .2 -> 84.4 ± .3 | Helped |
| Apply SSPL to ERM | ERM · PACS (Art Painting, Cartoon, Photo, Sketch) | accuracy: 75.8 ± .4 -> 79.3 ± .4; 72.7 ± .2 -> 75.3 ± .2; 94.9 ± .2 -> 95.9 ± .5; 62.9 ± .1 -> 69.5 ± .4 | Helped |
| Apply SSPL to ERM | ERM · Office-Home (Artistic, Clipart, Product, RealWorld) | accuracy: 58.6 ± .3 -> 59.4 ± .4; 47.9 ± .4 -> 49.2 ± .2; 73.7 ± .4 -> 74.3 ± .3; 75.8 ± .4 -> 76.0 ± .3 | Helped |
| Apply SSPL to ERM | ERM · VLCS (Caltech, Labelme, Pascal, Sun) | accuracy: 96.0 ± .3 -> 96.9 ± .3; 63.2 ± .4 -> 65.0 ± .2; 74.1 ± .2 -> 75.7 ± .4; 70.7 ± .6 -> 73.7 ± .4 | Helped |
| Replace SSPL with classic SPL (ablation) | ERM · Digits (MNIST, MNIST-M, SVHN, SYN) | accuracy: 96.4 ± .2 -> 90.5 ± .5; 62.6 ± .1 -> 55.4 ± .4; 66.7 ± .1 -> 48.7 ± .4; 83.8 ± .2 -> 64.9 ± .8 | Hurt |
| Replace SSPL with classic SPL (ablation) | ERM · PACS (Art Painting, Cartoon, Photo, Sketch) | accuracy: 75.8 ± .4 -> 66.4 ± .9; 72.7 ± .2 -> 71.0 ± .4; 94.9 ± .2 -> 91.0 ± .2; 62.9 ± .1 -> 55.3 ± .8 | Hurt |
| Replace SSPL with classic SPL (ablation) | ERM · Office-Home (Artistic, Clipart, Product, RealWorld) | accuracy: 58.6 ± .3 -> 48.0 ± .5; 47.9 ± .4 -> 38.5 ± .4; 73.7 ± .4 -> 60.0 ± .3; 75.8 ± .4 -> 60.1 ± .7 | Hurt |
| Replace SSPL with classic SPL (ablation) | ERM · VLCS (Caltech, Labelme, Pascal, Sun) | accuracy: 96.0 ± .3 -> 96.5 ± .4; 63.2 ± .4 -> 64.0 ± .3; 74.1 ± .2 -> 64.5 ± .5; 70.7 ± .6 -> 54.0 ± .5 | Mixed |
| Replace GDM with cross-entropy loss difficulty (ablation) | SSPL with ERM · Digits (MNIST, MNIST-M, SVHN, SYN) | accuracy: 96.6 ± .3 vs 96.9 ± .1; 64.7 ± .3 vs 65.5 ± .4; 67.0 ± .2 vs 67.7 ± .4; 84.1 ± .2 vs 84.4 ± .3 | Hurt |
| Replace GDM with cross-entropy loss difficulty (ablation) | SSPL with ERM · PACS | accuracy: 77.8 ± .2 vs 79.3 ± .4; 73.6 ± .3 vs 75.3 ± .2; 95.3 ± .3 vs 95.9 ± .5; 67.3 ± .2 vs 69.5 ± .4 | Hurt |
| Replace GDM with cross-entropy loss difficulty (ablation) | SSPL with ERM · Office-Home | accuracy: 59.2 ± .2 vs 59.4 ± .4; 48.5 ± .3 vs 49.2 ± .2; 74.1 ± .2 vs 74.3 ± .3; 75.5 ± .1 vs 76.0 ± .3 | Hurt |
| Replace GDM with cross-entropy loss difficulty (ablation) | SSPL with ERM · VLCS | accuracy: 96.6 ± .2 vs 96.9 ± .3; 64.3 ± .3 vs 65.0 ± .2; 75.0 ± .3 vs 75.7 ± .4; 72.9 ± .2 vs 73.7 ± .4 | Hurt |
Surprising findings
- Classic self-paced learning combined with ERM causes large accuracy drops on several domains (e.g., about 20% across Office-Home domains), while SSPL improves them.
Limitations
- Gains are reported on image classification benchmarks only; extension to other tasks is untested.
- Local (batch-wise) difficulty ranking is an approximation of the global ranking.
Experimental setup
Datasets: Digits PACS Office-Home VLCS
Benchmarks: Digits PACS Office-Home VLCS
Metrics: top-1 accuracy 95% confidence interval
Baselines: ERM CrossGrad MixStyle DomainMix EFDMix RandomErasing RandomRotation ColorJitter classic SPL
Builds on: Self-Paced Learning Curriculum Learning ERM
Closest papers by meaning
- Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · similarity 0.86)
- Self-Training with Dynamic Weighting for Robust Gradual Domain Adaptation (NeurIPS 2025 · similarity 0.86)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.86)
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.86)
- Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.85)
- Enhancing Evolving Domain Generalization through Dynamic Latent Representations (AAAI 2024 · 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.