Research map / Architectures and efficiency / RNNs and normalization
Unlocking Noise-Resistant Vision: Key Architectural Secrets for Robust Models Against Gaussian Noise
Bum Jun Kim, Makoto Kawano, Yusuke Iwasawa, Yutaka Matsuo
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 |
|---|---|---|---|
| use patch size 32 instead of 16 (ablation) | ViT-B augreg in1k 224 · ImageNet-1K Gaussian noise | RankDiff: -375 -> -506 | Improved |
| use native resolution 224 instead of 384 (pre-resize noise) (ablation) | ViT-B/16 augreg in1k · ImageNet-1K Gaussian noise | RankDiff: -152 -> -375 | Improved |
| add noise after resizing instead of before (counterfactual) (ablation) | supervised ViT pairs 224 vs 384 · consistency gap 384 minus 224 | mean consistency gap: -0.50 (34/63) -> +4.18 (62/63) | Mixed |
| use average pooling in stem instead of max pooling (ablation) | ResNet-50 · FGVC-Aircraft (noisy test) | accuracy with noise: 27.3 -> 43.1 | Improved |
| use average pooling in stem instead of max pooling (ablation) | ResNet-50 · Oxford-IIIT Pet (noisy test) | accuracy with noise: 41.8 -> 49.3 | Improved |
| use average pooling in stem instead of max pooling (ablation) | ResNet-50 · Caltech-UCSD Birds-200-2011 (noisy test) | accuracy with noise: 26.1 -> 26.8 | No clear effect |
| use nearest-neighbor pooling instead of max pooling (ablation) | ResNet-50 · FGVC-Aircraft (noisy test) | accuracy with noise: 27.3 -> 24.7 | Hurt |
| use average-pooling downsampling (ResNet-T) instead of strided convolution (ResNet-S) (ablation) | ResNet-50 · Oxford-IIIT Pet (noisy test) | accuracy with noise: 22.9 -> 38.1 | Improved |
| replace OPENAI normalization constants with INCEPTION constants (ablation) | CLIP ViT-B/16 openai ft in12k · Oxford-IIIT Pet (noisy test) | accuracy with noise: 76.3 -> 87.2 | Improved |
| replace OPENAI normalization constants with IMAGENET constants (ablation) | CLIP ViT-B/16 openai ft in12k · Oxford-IIIT Pet (noisy test) | accuracy with noise: 76.3 -> 72.7 | Hurt |
| use INCEPTION constants instead of OPENAI (ablation) | supervised ViT-B/16 augreg in1k · Oxford-IIIT Pet (noisy test) | accuracy with noise: 88.2 -> 89.2 | No clear effect |
| use supervised ViT instead of CLIP ViT | ViT-B/16 224 · ImageNet-1K Gaussian noise | RankDiff: +74 -> -65 | Improved |
Closest papers by meaning
- VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models (ICML 2026 · similarity 0.88)
- Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning (ICML 2024 · similarity 0.86)
- Instant Video Models: Universal Adapters for Stabilizing Image-Based Networks (NeurIPS 2025 · similarity 0.86)
- LAION-C: An Out-of-Distribution Benchmark for Web-Scale Vision Models (ICML 2025 · similarity 0.86)
- Enhance Vision-Language Alignment with Noise (AAAI 2025 · similarity 0.86)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.86)
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.
