Learning Fractals by Gradient Descent

Authors: Cheng-Hao Tu, Hong-You Chen, David Carlyn, Wei-Lun Chao

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct two experiments to validate our approach.
Researcher Affiliation Academia Department of Computer Science and Engineering, The Ohio State University {tu.343, chen.9301, carlyn.1, chao.209}@osu.edu
Pseudocode Yes Algorithm 1: IFS generation process. See section 6 for details. Algorithm 2: IFS generation process via the FE layer. See subsection 6 for details.
Open Source Code Yes The code is provided at https://github.com/andytu28/Learning Fractals.
Open Datasets Yes We first reconstruct random fractal images generated following Fractal DB (Kataoka et al. 2020; Anderson and Farrell 2022). We then consider images that are not generated by fractals, including MNIST (hand-written digits) (Le Cun et al. 1998), FMNIST (fashion clothing) (Xiao, Rasul, and Vollgraf 2017), and KMNIST (hand-written characters) (Clanuwat et al. 2018).
Dataset Splits No For the image reconstruction task: 'As we would recover any given target, no training and test splits are considered.' For the GAN extension, it mentions 'training/test splits' but does not specify a separate validation set or its characteristics.
Hardware Specification No The paper mentions 'computational resources by the Ohio Supercomputer Center and AWS Cloud Credits for Research' but does not specify particular GPU or CPU models, or exact hardware configurations.
Software Dependencies No The paper mentions 'deep learning frameworks like Py Torch (Paszke et al. 2019) and Tensor Flow (Abadi et al. 2016)' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Each S is learned with a batch size 50 and a learning rate 0.05 for 1K SGD steps. [...] we set τ = 1 (RBF kernel bandwidth) and apply an Adam optimizer (Kingma and Ba 2015). The number of transformations in S is fixed to N = 10, and we set T = 300 per image.