Denoising and Regularization via Exploiting the Structural Bias of Convolutional Generators

Authors: Reinhard Heckel and Mahdi Soltanolkotabi

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We depict this phenomena in Figure 1 where we fit a randomly initialized over-parameterized convolutional generator to an image via running gradient descent on the objective L(C) = G(C) y 2 2. Here, G(C) is the convolutional generator with weight parameters C, and y is either a noisy image, a clean image, or noise. This experiment demonstrates that an over-parameterized convolutional network fits a natural image (Figure 1b) much faster than noise (Figure 1c).
Researcher Affiliation Academia Reinhard Heckel Dept. of Electrical and Computer Engineering Technical University of Munich reinhard.heckel@tum.de Mahdi Soltanolkotabi Dept. of Electrical and Computer Engineering University of Southern California soltanol@usc.edu
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes CODE AND ACKNOWLEDGEMENTS Code to reproduce the experiments is available at https://github.com/MLI-lab/ overparameterized_convolutional_generators.
Open Datasets Yes We compared the performance of those four methods on denoising 100 randomly chosen images from the Image Net validation set.
Dataset Splits No The paper describes fitting a single noisy image rather than traditional training/validation/test dataset splits. Appendix A describes evaluation on 100 images from the ImageNet validation set, which serve as evaluation data, not for traditional splitting for a supervised learning setup.
Hardware Specification No R. Heckel is partially supported by NSF award IIS-1816986, acknowledges support of the NVIDIA Corporation in form of a GPU, and would like to thank Tobit Klug for proofreading a previous version of this manuscript. The mention of 'a GPU' is too general and does not provide specific model or configuration details.
Software Dependencies No The paper mentions "pytorch and tensorflow" as standard implementations for upsampling operations, but does not provide specific version numbers for these or any other software dependencies crucial for replication.
Experiment Setup Yes early stopped after 1900 iterations, again with the same stopping time as proposed in the original paper, iii) an under-parameterized deep decoder with five layers and k = 128 channels in each layers, trained for 3000 iterations (which is close to convergence), iv) an over-parameterized deep decoder with five layers and k = 512 channels in each layer, early stopped at 1500 iterations.