Reference library
FractalNet: Ultra-Deep Neural Networks without Residuals
Paper summary
FractalNet recursively expands a feedforward convolutional graph so short and long transformed paths join by elementwise means. Local/global drop-path makes individual columns learn usable predictors. On CIFAR-100 without augmentation the unregularized full model reports35.34% error and the regularized model28.20%; the regularized configuration also uses dropout. On ImageNet the34-layer model reports24.12% top-1 error against24.19% for the cited34-layer ResNet configuration. A plain40-layer CIFAR-100++ network reports38.84% error, a40-layer column trained inside the fractal graph31.75%, and the full graph27.40%. These support trainability and a multiscale training-scaffold interpretation. They do not isolate a universal benefit of self-similarity. Increasing depth from80 to160 raises error from26.46% to27.38% in the separate reduced-width comparison.
Relevance to the project
Assistant interpretation: a concrete recursive graph-construction candidate for a future controlled comparison. It does not establish sustained activity, plastic stability, useful recurrent dynamics or effective crossover in the adaptive F/H system. Preserve the existing architecture and initialization; compare path-length distribution, degree, parameters and total training/evaluation cost before proposing adoption.
Image classification with supervised gradient training, BN, ReLU, augmentation and regularization. Equal depth does not establish equal parameter/FLOP budget. Tables3–4 use settings not directly comparable with Table1. No reproduction here; the author README supplies local drop-path but explicitly lacks the full local+global release and augmentation implementation.
Source record
- Reference ID
- fractalnet-2017
- Archived version
- 1605.07648v4
- Summary provenance
- reference-library/papers/fractalnet-2017/metadata.json · summary
- Review status
- main-text reading and selected visual verification; not replication
- Reading scope
- Main text PDF pages1–9, including methods and Tables1–4; page7 visually inspected against rendered original. Pages10–11 bibliography not fully read. Author repository README.rst read, not the complete implementation and no code executed.