Research map / Label-efficient and robust learning / Domain adaptation and generalization
Rethinking Multi-domain Generalization with A General Learning Objective
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 GMDG, a general multi-domain generalization objective that relaxes the static P(Y|D) assumption with a learnable Y-mapping and combines domain-invariant conditional feature learning, posterior maximization, prior regularization, and invalid-causality suppression.
Problem
Multi-domain generalization methods typically assume a static target marginal P(Y|D) across domains, which fails for regression and segmentation, and existing objectives lack a general formulation.
Approach
Learnable mappings phi (for X) and psi (for Y) project inputs and targets into a shared latent space where domain-independent conditional features are learned, yielding a general four-term objective (two aims, two regularizers) that interprets and subsumes prior mDG methods; a conditional feature shift term suppresses invalid causality.
Key results
- On DomainBed with a RegNetY-16GF oracle, GMDG averages 75.1% vs 74.1% for MIRO without SWAD, and 78.2% vs 77.3% with SWAD.
- With a single pre-trained oracle, GMDG surpasses SIMPLE++ (283-model ensemble) by about 0.1% average accuracy in classification.
- On NYU Depth V2 leave-one-domain-out depth estimation, GMDG reduces average SILog from 12.9258 (Swin-L backbone) to 12.2514.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the full GMDG objective to monocular depth estimation | VA-DepthNet with Swin-L · NYU Depth V2 Avg. | Abs Rel: 13.74 -> 13.27 | Helped |
| Add the full GMDG objective to monocular depth estimation | VA-DepthNet with Swin-L · NYU Depth V2 Avg. | δ1: 81.96 -> 82.52 | Helped |
| Add the full GMDG objective | DeepLabv3+ segmentation · GTAV-trained mDG; average of Cityscapes, BDD-100K, and Mapillary | mIoU: 37.48 -> 37.65 | Helped |
| Use GMDG instead of MIRO with RegNetY-16GF oracle | Domain generalization classification · TerraIncognita | accuracy: 58.9±1.3 -> 60.7±1.8 | Helped |
| Use GMDG instead of MIRO with RegNetY-16GF oracle | Domain generalization classification · PACS | accuracy: 97.4±0.2 -> 97.3±0.1 | Hurt |
| Add GAim1 to GAim2+GReg1 (ablation) | Classification with oracle prior · HomeOffice | average accuracy: 84.0±0.2 -> 84.5±0.2 | Helped |
| Add GReg2 to GAim2+GReg1 (ablation) | Classification with oracle prior · HomeOffice | average accuracy: 84.0±0.2 -> 84.0±0.2 | No clear effect |
| Add iReg2 instead of GReg2 to GAim2+GReg1 (ablation) | Classification with oracle prior · HomeOffice | average accuracy: 84.0±0.2 -> 83.9±0.1 | Hurt |
| Add iAim1 to GAim2+GReg1 (ablation) | Classification with oracle prior · HomeOffice | average accuracy: 84.0±0.2 -> 84.1±0.2 | No clear effect |
| Add GReg2 to GAim2 without prior (ablation) | Classification without oracle prior · HomeOffice | average accuracy: 79.6±0.2 -> 80.3±0.3 | Helped |
| Add GAim1 to GAim2 without prior (ablation) | Classification without oracle prior · HomeOffice | average accuracy: 79.6±0.2 -> 79.9±0.4 | Helped |
| Add the Y-mapping to LA1 (ablation) | Toy regression with synthetic data · Synthetic data, affine transformations without DCDS | MSE: 0.3537 -> 0.3369 | Helped |
Surprising findings
- Combining the prior-based MIRO-style objective with unconditional covariance shift regularization (CORAL-style) degrades performance when a large pre-trained oracle is used.
- Previous methods such as VA-DepthNet perform worse than their backbone on depth estimation, since they lack an mDG-targeted objective.
Limitations
- Assumes equal domain weights when minimizing the generalized Jensen-Shannon divergence, not handling domain imbalance.
- Gains over strong baselines are often small (e.g., about 0.1-0.3% on DomainBed with ResNet-50), and the paper relies on a pre-trained oracle model.
Experimental setup
Datasets: NYU Depth V2 GTAV SYNTHIA Cityscapes PACS VLCS OfficeHome TerraIncognita DomainNet BDD-100K Mapillary
Benchmarks: NYU Depth V2 Cityscapes BDD-100K Mapillary GTAV SYNTHIA PACS VLCS OfficeHome TerraIncognita DomainNet
Metrics: SILog Abs Rel RMS Sq Rel RMS log delta1 delta2 mIoU accuracy
Baselines: ERM DANN CDANN CIDG CORAL MDA MIRO SIMPLE SIMPLE++ RobustNet VA-DepthNet DeepLabv3+
Builds on: MIRO MDA RobustNet CDANN DANN CORAL
Closest papers by meaning
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.91)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.89)
- Is Large-scale Pretraining the Secret to Good Domain Generalization? (ICLR 2025 · similarity 0.89)
- One-Step Generalization Ratio Guided Optimization for Domain Generalization (ICML 2025 · similarity 0.89)
- Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.88)
- Towards Generalizing to Unseen Domains with Few Labels (CVPR 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.