Research map / Optimization / Optimizers
Learning Fractals by Gradient Descent
Cheng-Hao Tu, Hong-You Chen, David Carlyn, Wei-Lun Chao
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 |
|---|---|---|---|
| Optimize fractal parameters with gradient descent rather than predict them with an inverse encoder | IFS fractal reconstruction · FractalDB, MNIST, KMNIST, and FMNIST | MSE: FractalDB: 0.3889 ± 0.0000 -> 0.0617±0.0009; MNIST: 0.1724 ± 0.0000 -> 0.0202 ± 0.0005; KMNIST: 0.3016 ± 0.0000 -> 0.078 | Improved |
| Use the Expectation rather than Fixed objective (ablation) | IFS fractal reconstruction · FractalDB, MNIST, KMNIST, and FMNIST | MSE: FractalDB: 0.0752±0.0027 -> 0.0630±0.0008; MNIST: 0.0312 ± 0.0002 -> 0.0223 ± 0.0004; KMNIST: 0.0979 ± 0.0008 -> 0.0826 | Improved |
| Clamp rendered pixels to [0, 1] (ablation) | IFS fractal reconstruction with the Fixed objective and no noisy SGD · FractalDB, MNIST, KMNIST, and FMNIST | MSE: FractalDB: 0.0790±0.0006 -> 0.0752±0.0027; MNIST: 0.0475 ± 0.0012 -> 0.0312 ± 0.0002; KMNIST: 0.1023 ± 0.0016 -> 0.0979 | Mixed |
| Add noisy SGD (ablation) | IFS fractal reconstruction with the Expectation objective and clamping · FractalDB, MNIST, KMNIST, and FMNIST | MSE: FractalDB: 0.0630±0.0008 -> 0.0617±0.0009; MNIST: 0.0223 ± 0.0004 -> 0.0202 ± 0.0005; KMNIST: 0.0826 ± 0.0010 -> 0.0781 | Improved |
| Remove affine-matrix reparameterization (ablation) | IFS fractal reconstruction · FractalDB, MNIST, KMNIST, and FMNIST | MSE | Hurt |
| Increase the number of transformations from 2 to 6 (ablation) | IFS fractal reconstruction · FractalDB and MNIST | MSE: FractalDB: 0.0954±0.0035 -> 0.0687±0.0006; MNIST: 0.0423±0.0013 -> 0.0237±0.0015 | Improved |
| Increase the number of transformations from 6 to 10 (ablation) | IFS fractal reconstruction · FractalDB and MNIST | MSE: FractalDB: 0.0687±0.0006 -> 0.0627±0.0013; MNIST: 0.0237±0.0015 -> 0.0212±0.0009 | Improved |
| Replace Adam with SGD (ablation) | IFS fractal reconstruction · MNIST | MSE: 0.0212±0.0009 -> 0.0656±0.0014 | Hurt |
| Train fractal parameters with a GAN loss to match target image distributions | A set of 100 IFS fractals used for pre-training · MNIST, FMNIST, and KMNIST | test accuracy | Improved |
Closest papers by meaning
- Pre-Training Without Natural Images (IJCV 2022 · similarity 0.86)
- Encoding Spatial Distribution of Convolutional Features for Texture Representation (NeurIPS 2021 · similarity 0.86)
- Self-Similarity Priors: Neural Collages as Differentiable Fractal Representations (NeurIPS 2022 · similarity 0.85)
- Fractal Calibration for Long-tailed Object Detection (CVPR 2025 · similarity 0.81)
- FRN: Fractal-Based Recursive Spectral Reconstruction Network (NeurIPS 2025 · similarity 0.81)
- On the Limitations of Fractal Dimension as a Measure of Generalization (NeurIPS 2024 · similarity 0.80)
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.
