Research map / Label-efficient and robust learning / Domain adaptation and generalization
Characterizing and Avoiding Negative Transfer
Check 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 the discriminator gate to deep transfer methods | DAN, DCORAL, DANN, ADDA, PADA, and GTA · Four benchmarks, eight transfer tasks | accuracy: ↑8.3; ↑5.2; ↑8.1; ↑7.1; ↑7.5; ↑13.3; ↑8.9; ↑10.4 | Helped |
| Add the discriminator gate to DANN | DANN · W→D and A→D; ε=0.0 | accuracy: ↓0.2; ↓0.2 | No clear effect |
| Add the discriminator gate to DANN | DANN · W→D and A→D; ε=0.9, L%=30% | accuracy: ↑9.4; ↑29.3 | Helped |
| Add the discriminator gate to DANN | DANN · A→D, ε=0.7 and 0.3; L%=30% and 10% | accuracy: 70.4 -> 82.5; 49.4 -> 58.7; 72.5 -> 82.7; 54.3 -> 60.7 | Helped |
| Add a discriminator gate, without joint-label distribution matching (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 66.0 | Hurt |
| Use label information but remove the discriminator gate (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 64.9 | Hurt |
| Match only the joint distribution (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 70.6 | Hurt |
| Match only the marginal distribution (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 69.3 | Hurt |
| Remove marginal and joint distribution matching (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 67.3 | Hurt |
| Exclude perturbed source data using a human oracle (ablation) | DANN_gate · A→D, average across four settings | accuracy: 71.2 -> 78.7 | Hurt |
| Increase source-target distribution divergence | DANN · Office-31 W→D and A→D | negative transfer gap | Hurt |
| Increase the amount of labeled target data | DANN relative to its target-only baseline · Office-31 transfer tasks | negative transfer gap | Hurt |
Closest papers by meaning
- Residual Feature Integration is Sufficient to Prevent Negative Transfer (ICLR 2026 · similarity 0.89)
- Adversarial Training Helps Transfer Learning via Better Representations (NeurIPS 2021 · similarity 0.88)
- Adversarially-Trained Deep Nets Transfer Better: Illustration on Image Classification (ICLR 2021 · similarity 0.87)
- Mitigating Negative Transfer in Multi-Task Learning with Exponential Moving Average Loss Weighting Strategies (Student Abstract) (AAAI 2023 · similarity 0.87)
- A Data-Based Perspective on Transfer Learning (CVPR 2023 · similarity 0.87)
- Initialization Matters for Adversarial Transfer Learning (CVPR 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.