Research map / Architectures and efficiency / RNNs and normalization
Diversity-aware Weight Perturbation Promotes Robust Adaptation
Zibo Chen, Ruxin Li, Zilu Wang
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 diversity-aware affinity weighting on top of plain weight noise injection (ablation) | DWP (ResNet) · CIFAR10 (mismatch 0.7) | accuracy: additional 14.0% accuracy boost | Improved |
| add plain multiplicative weight-noise injection during training (ablation) | baseline ResNet · CIFAR10 (mismatch 0.7) | accuracy: 12.0% to 50.0% | Improved |
| increase noise realizations n from 5 to 10 (ablation) | DWP (ResNet) · CIFAR10 | accuracy: 60.01 -> 65.61 | Improved |
| increase noise realizations n from 10 to 20 (ablation) | DWP (ResNet) · CIFAR10 | accuracy: 65.61 -> 66.43; training time 32.32s -> 60.46s | No clear effect |
| match training mismatch level to test level (0.3 vs 0.1) (ablation) | DWP (ResNet) · CIFAR10 (test mismatch 0.3) | accuracy: 70.33 -> 87.54 | Improved |
| raise training mismatch level beyond test level (0.3 to 0.5) (ablation) | DWP (ResNet) · CIFAR10 (test mismatch 0.3) | accuracy: 87.54 -> 75.59 | Hurt |
| train with DWP instead of standard training | F-MNIST CNN · F-MNIST (mismatch 0.7) | accuracy: 53.95 -> 90.01 | Improved |
| train with DWP instead of standard training | CIFAR10 ResNet · CIFAR10 (mismatch 0.5) | accuracy: 18.02 -> 73.00 | Improved |
| train with DWP instead of standard training | CIFAR10 ViT · CIFAR10 (mismatch 0.7) | accuracy: 20.15 -> 65.30 | Improved |
| train with DWP instead of Forward Noise at low mismatch | F-MNIST CNN · F-MNIST (mismatch 0.1) | accuracy: 91.20 vs 91.74 | No clear effect |
| apply DWP to Qwen3-1.7B classification (clean accuracy cost) | Qwen3-1.7B · iFLYTEK 119 (noise 0.2) | clean accuracy: 58.06 -> 53.56 | Hurt |
| apply DWP to Qwen3-1.7B under strong noise | Qwen3-1.7B · iFLYTEK 119 (noise 0.4) | noisy accuracy: 1.26 -> 8.93 | Improved |
Closest papers by meaning
- AdMiT: Adaptive Multi-Source Tuning in Dynamic Environments (CVPR 2025 · similarity 0.85)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (NeurIPS 2025 · similarity 0.85)
- Rethinking Weight Decay for Robust Fine-Tuning of Foundation Models (NeurIPS 2024 · similarity 0.84)
- $\boldsymbol{\mu}\mathbf{P^2}$: Effective Sharpness Aware Minimization Requires Layerwise Perturbation Scaling (NeurIPS 2024 · similarity 0.84)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.84)
- Power Lines: Scaling laws for weight decay and batch size in LLM pre-training (NeurIPS 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.
