Research map / Label-efficient and robust learning / Recognition and long-tail
Pre-Training Without Natural Images
Hirokatsu Kataoka, Kazushige Okayasu, Asato Matsumoto, Eisuke Yamagata, Ryosuke Yamada, Nakamasa Inoue, Akio Nakamura, Yutaka Satoh
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 |
|---|---|---|---|
| render fractals with 3x3 patches instead of 1x1 points (ablation) | FractalDB · CIFAR-10 | accuracy (%): 87.4 -> 92.1 | Improved |
| render fractals with 3x3 patches instead of 1x1 points (ablation) | FractalDB · CIFAR-100 | accuracy (%): 66.1 -> 72.0 | Improved |
| use fixed patch instead of random patch pattern (ablation) | FractalDB · CIFAR-100 | accuracy (%): 72.0 -> 73.6 | Improved |
| filling rate 0.05 to 0.15 (ablation) | FractalDB · CIFAR-10 | accuracy (%): 91.8, 92.0, 91.7 | No clear effect |
| raise filling rate to 0.25 (ablation) | FractalDB · CIFAR-100 | accuracy (%): 72.3 -> 63.2 | Hurt |
| widen intra-category weight interval from 0.1 to 0.4 (ablation) | FractalDB · CIFAR-10 | accuracy (%): 92.1 -> 92.7 | Improved |
| widen intra-category weight interval to 0.5 (ablation) | FractalDB · CIFAR-10 | accuracy (%): 92.7 -> 91.8 | Hurt |
| increase image size from 256 to 1024 (ablation) | FractalDB · CIFAR-100 | accuracy (%): 73.6 -> 68.6 | Hurt |
| increase number of dots from 100k to 400k (ablation) | FractalDB · CIFAR-10 | accuracy (%): 91.3 -> 90.4 | Hurt |
| assign categories by DeepCluster k-means instead of IFS parameters (ablation) | FractalDB · CIFAR-100 | accuracy (%): 77.3 vs 57.0 | Hurt |
| freeze conv1 during fine-tuning (ablation) | FractalDB-1k · CIFAR-100 | accuracy (%): 75.7 -> 72.2 | Hurt |
| freeze conv1-5 during fine-tuning (ablation) | FractalDB-1k · CIFAR-10 | accuracy (%): 93.4 -> 49.4 | Hurt |
Closest papers by meaning
- Can Vision Transformers Learn without Natural Images? (AAAI 2022 · similarity 0.88)
- Pre-training Vision Transformers with Very Limited Synthesized Images (ICCV 2023 · similarity 0.87)
- Replacing Labeled Real-Image Datasets With Auto-Generated Contours (CVPR 2022 · similarity 0.87)
- Learning Fractals by Gradient Descent (AAAI 2023 · similarity 0.86)
- Pre-training Vision Models with Mandelbulb Variations (CVPR 2024 · similarity 0.83)
- Non-Natural Image Understanding with Advancing Frequency-based Vision Encoders (CVPR 2025 · similarity 0.83)
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.
