Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain Separation Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 private-shared subspace separation with difference (orthogonality) loss to DSN w/ DANN (ablation) | DSN w/ DANN · MNIST to MNIST-M, Synth Digits to SVHN, SVHN to MNIST, Synth Signs to GTSRB | classification accuracy (%): +2.97, +2.01, +2.24, +1.12 points vs. beta=0 | Helped |
| replace scale-invariant MSE reconstruction loss with standard L2 MSE (ablation) | DSN w/ DANN · MNIST to MNIST-M, Synth Digits to SVHN, SVHN to MNIST, Synth Signs to GTSRB | classification accuracy (%): -2.81, -2.24, -3.33, -0.90 points | Hurt |
| use DANN as similarity loss in DSN instead of MMD | DSN · all four digit/sign scenarios | classification accuracy (%): e.g. MNIST-M 83.2 vs 80.5; SVHN to MNIST 82.7 vs 72.2 | Helped |
| private-shared separation (DSN) with DANN similarity vs DANN alone | DANN · MNIST to MNIST-M, Synth Digits to SVHN, SVHN to MNIST, Synth Signs to GTSRB | classification accuracy (%): +5.8, +0.9, +12.0, +0.2 points | Helped |
| private-shared separation (DSN) with DANN on LINEMOD pose estimation | DANN · Synth Objects to LINEMOD | mean angle error (degrees): 56.58 -> 53.27 (lower is better) | Helped |
| private-shared separation with MMD similarity vs plain MMD regularization | MMD · MNIST to MNIST-M, Synth Digits to SVHN, SVHN to MNIST, Synth Signs to GTSRB | classification accuracy (%): +3.6, +0.5, +1.1, +1.5 points | Helped |
| private-shared separation with MMD similarity on LINEMOD | MMD · Synth Objects to LINEMOD | classification accuracy (%): 72.35 -> 99.72 | Helped |
| replace MMD similarity with CORAL-style alignment (CORAL baseline) vs source-only | CORAL · Synth Digits to SVHN | classification accuracy (%): 86.7 -> 85.2 | Hurt |
| Correlation Regularization (CorReg) similarity loss used as a new domain adaptation method | CORAL (baseline comparison) · MNIST to MNIST-M, Synth Digits to SVHN, SVHN to MNIST, Synth Signs to GTSRB | classification accuracy (%): 62.06 vs 57.7; 87.33 vs 85.2; 69.20 vs 63.1; 90.75 vs 86.9 | Helped |
Closest papers by meaning
- Domain Adaptation with Invariant Representation Learning: What Transformations to Learn? (NeurIPS 2021 · similarity 0.91)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.89)
- Adaptive Methods for Real-World Domain Generalization (CVPR 2021 · similarity 0.89)
- Visual Domain Adaptation by Consensus-Based Transfer to Intermediate Domain (AAAI 2020 · similarity 0.89)
- On Learning Invariant Representations for Domain Adaptation (ICML 2019 · similarity 0.88)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.87)
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.