Research map / Label-efficient and robust learning / Domain adaptation and generalization
Interpretations of Domain Adaptations via Layer Variational Analysis
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 |
|---|---|---|---|
| analytic last-layer update (LVA) instead of gradient descent finetuning | BLSTM speech enhancement, mismatched speakers · Speech enhancement mismatched speakers | PESQ average: 2.16 vs 1.92 (GD) | Helped |
| analytic last-layer update (LVA) instead of gradient descent finetuning | FullSubNet speech enhancement adaptation · Speech enhancement FullSubNet | PESQ average: 2.94 vs 2.82 (GD) | Helped |
| analytic last-layer update (LVA) instead of gradient descent finetuning | FullSubNet speech enhancement adaptation · Speech enhancement FullSubNet | STOI average: 0.94 vs 0.94 (GD) | No clear effect |
| analytic last-layer update (LVA) instead of gradient descent finetuning | BLSTM speech enhancement, mismatched speakers · Speech enhancement mismatched speakers | STOI average: 0.85 vs 0.84 (GD) | Helped |
| finetuning the last layer (GD or LVA) vs no finetuning (ablation) | FullSubNet speech enhancement adaptation · Speech enhancement FullSubNet | PESQ average: 2.82 (GD), 2.94 (LVA) vs 2.28 (no finetune) | Helped |
| LVA with optimal transport domain alignment vs LVA without alignment (ablation) | BLSTM speech enhancement, noise/SNR adaptation · Speech enhancement noise/SNR adaptation | L2 loss / PESQ: LVA-OT achieves similar performance to LVA | No clear effect |
| analytic LVA update instead of GD, with few adaptation samples | BLSTM speech enhancement, noise/SNR adaptation · Speech enhancement noise/SNR adaptation | L2 loss: LVA significantly outperformed GD when target samples were fewer | Helped |
| analytic one-step LVA update instead of 12000-epoch GD on last layer | 1D time series regression · Synthetic 1D signal regression | L2 loss: LVA reached lowest L2 loss vs GD (values only in figure) | Helped |
| LVA analytic finetuning of last layer instead of GD for SRCNN deblurring | SRCNN image deblurring · SET14 deblurring adaptation | PSNR: LVA highest PSNR at N=256 vs GD; GD with N=16384 still not compatible | Helped |
| finetuning last layer via LVA vs GD in terms of theory: LVA as theoretical limit | Speech enhancement · Speech enhancement | L2 loss: LVA L2 loss notably less than GD | Helped |
Closest papers by meaning
- Residual Parameter Transfer for Deep Domain Adaptation (CVPR 2018 · similarity 0.90)
- Unsupervised Domain Adaptation with Residual Transfer Networks (NeurIPS 2016 · similarity 0.89)
- Domain Adaptation with Invariant Representation Learning: What Transformations to Learn? (NeurIPS 2021 · similarity 0.88)
- On Learning Domain-Invariant Representations for Transfer Learning with Multiple Sources (NeurIPS 2021 · similarity 0.88)
- Transferable Adversarial Training: A General Approach to Adapting Deep Classifiers (ICML 2019 · similarity 0.88)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 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.