Reference library
Pre-training without Natural Images
Paper summary
FractalDB generates labeled synthetic images using iterated affine transformations and parameter-defined categories, then pretrains a conventional ResNet-50 before natural-image fine-tuning. It uses fractals as data, not network wiring. Table6 reports CIFAR-100 accuracy62.7% from scratch,75.7% after FractalDB-1k,77.3% after FractalDB-10k and84.6% after ImageNet-1k pretraining. FractalDB-10k reaches50.8% on Places365 against50.3% for ImageNet pretraining but does not win all tasks. Table9 compares roughly million-image generated datasets: CIFAR-10075.7% for FractalDB-1k against71.1% for Perlin-1296 and68.1% for Bezier-1024. Category labels, rendering, filling rate and augmentations matter; larger image size/dot count are not uniformly better. Synthetic pretraining incurs substantial generation and optimization cost.
Relevance to the project
Assistant interpretation: useful if a separate visual encoder or formula-generated diagnostic task is later needed. Current scalar/vector recurrent F/H controllers do not acquire this advantage merely by introducing fractal connectivity. A synthetic training curriculum is a separate proposal, requiring task relevance and complete-cost comparisons.
Read the January2021 arXiv preprint, not the extended2022 IJCV version advertised by the current author repository. Numerical comparisons belong to this preprint. Common optimization settings do not equalize all dataset sizes, compute or tuning effort. Natural-image baselines outperform on important tasks. Category count, rendering and curriculum can explain gains without a unique fractality mechanism; table/prose inconsistencies are not silently repaired.
Source record
- Reference ID
- fractaldb-2021
- Archived version
- 2101.08515v1
- Summary provenance
- reference-library/papers/fractaldb-2021/metadata.json · summary
- Review status
- main-text reading and selected visual verification; not replication
- Reading scope
- PDF pages1–10 through conclusion and start of references, Tables1–10; page8 visually inspected. Bibliography pages11–12 not fully read. Current author implementation README read in full; its2022 updates are not treated as the exact2021 implementation. No training or replication.