Research map / Optimization / Implicit bias and generalization
LDT: Layer-Decomposition Training Makes Networks More Generalizable
Zaizuo Tang, Zongqi Yang, Yu-Bin Yang
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 |
|---|---|---|---|
| add Layer-Decomposition Training (without DPU) (ablation) | MambaIR · DRealSR SR, Olympus source, Canon target | PSNR: 30.93 -> 32.33 | Improved |
| add Layer-Decomposition Training (without DPU) (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 30.81 -> 31.25 | Improved |
| add Layer-Decomposition Training (without DPU) (ablation) | MambaIR · DRealSR SR, Olympus source, Pan target | SSIM: 0.8688 -> 0.8631 | Hurt |
| add Dynamic Parameter Update to LDT (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 31.25 -> 32.15 | Improved |
| add Dynamic Parameter Update to LDT (ablation) | MambaIR · DRealSR SR, Olympus source, Canon target | PSNR: 32.33 -> 32.80 | Improved |
| partition layers by gradient variance instead of randomly (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 30.88 -> 32.15 | Improved |
| partition layers by gradient variance instead of mean gradient magnitude (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 31.86 -> 32.15 | Improved |
| partition by normalized variance (Var/Mean) instead of variance (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 32.15 -> 31.89 | No clear effect |
| reassign stable/unstable layers dynamically every 5,000 steps instead of static partition (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 32.15 -> 31.85 | Mixed |
| project variance rank into update coefficients instead of normalized variance (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 31.01 -> 32.15 | Improved |
| use single-branch LDT-S instead of dual-branch LDT (efficiency) (ablation) | MambaIR · Training memory, 48x48 patches | training memory (GB): 20.25 -> 15.21 | Improved |
| use single-branch LDT-S instead of dual-branch LDT (ablation) | MambaIR · DRealSR SR, Olympus source, Sony target | PSNR: 32.15 -> 31.46 | Hurt |
Closest papers by meaning
- Neural Network Optimization Reimagined: Decoupled Techniques for Scratch and Fine-Tuning (TPAMI 2026 · similarity 0.86)
- The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes (NeurIPS 2022 · similarity 0.85)
- LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging (ICLR 2025 · similarity 0.85)
- Stable Coresets via Posterior Sampling: Aligning Induced and Full Loss Landscapes (NeurIPS 2025 · similarity 0.84)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.84)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.84)
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.
