Research map / Label-efficient and robust learning / Domain adaptation and generalization
Loss Function Learning for Domain Generalization by Implicit Gradient
Boyan Gao, Henry Gouk, Yongxin Yang, Timothy Hospedales
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 cross-entropy with learned ITL loss in ERM | ERM on PACS (ResNet18) · PACS | accuracy (%): +5.5 over ERM+CE (78.9 to 84.4) | Improved |
| replace cross-entropy with learned ITL loss in ERM | ERM on DomainBed (ResNet50) · DomainBed multi-source average (VLCS, PACS, TerraInc, OfficeHome, DomainNet) | accuracy (%): improves on ERM on all listed datasets | Improved |
| replace cross-entropy with learned ITL loss in ERM | ERM on single-source DomainBed · Single-source DomainBed (PACS, TerraInc, OfficeHome, VLCS) | accuracy (%): PACS 51.85 to 56.54; TerraInc 32.13 to 35.09 | Improved |
| replace cross-entropy with ITL loss in RSC (ablation) | RSC on PACS (ResNet18) · PACS | accuracy (%): 81.9 to 82.7 (+0.8) | Improved |
| compare ITL to other manually designed robust losses in ERM (ablation) | ERM on DomainBed PACS (ResNet50) · DomainBed PACS | accuracy (%): ITL 86.4 vs Focal 84.6, LS 84.9, SCE 84.2, GCE 83.0 | Improved |
| learn SCE linear-combination weights with IFT meta-learning (iSCE) instead of Taylor loss (ablation) | ERM on PACS (ResNet18) · PACS | accuracy (%): iSCE 82.1 vs SCE 80.3 (improves) but below ITL 84.4 | Mixed |
| meta-train loss on RotatedKMNIST instead of RotatedMNIST (ablation) | ITL meta-training · OfficeHome | accuracy (%): 68.74 vs 69.03 (within std) | No clear effect |
| learn loss per task on ColoredMNIST instead of a generic loss (ablation) | ITL on ColoredMNIST · ColoredMNIST | accuracy (%): 52.2 vs 52.0 (slight) | No clear effect |
| use implicit gradient (IFT) instead of CMA-ES for meta-training the loss (ablation) | ITL meta-training · meta-train loss learning | outer loop accuracy: smoother convergence to higher accuracy; K times fewer model trainings | Improved |
| ITL loss in-domain vs cross-domain PACS testing (ablation) | ITL vs CE on PACS · PACS in-domain test accuracy | accuracy (%): 98.96 vs 99.34 in-domain; cross-domain 84.4 vs 81.5 | Hurt |
| replace CE with ITL in ERM | ERM on DomainBed ColoredMNIST · ColoredMNIST | accuracy (%): 52.0 vs 51.5 (+0.5) | No clear effect |
Closest papers by meaning
- Lost Domain Generalization Is a Natural Consequence of Lack of Training Domains (AAAI 2024 · similarity 0.89)
- Understanding Domain Generalization: A Noise Robustness Perspective (ICLR 2024 · similarity 0.88)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.88)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.88)
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.88)
- Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.88)
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.
