Super-resolution with deep convolutional sufficient statistics

Authors: Joan Bruna, Pablo Sprechmann, Yann Lecun

ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate experimentally the proposed framework in the image super-resolution task... and 4 EXPERIMENTAL EVALUATION
Researcher Affiliation Collaboration Joan Bruna Department of Statistics University of California, Berkeley joan.bruna@berkeley.edu Pablo Sprechmann Courant Institute New York University pablo@cims.nyu.edu Yann Le Cun Facebook, Inc., & Courant Institute New York University yann@cims.nyu.edu
Pseudocode No The paper describes the inference process in text but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about releasing source code for the described methodology, nor does it provide any links to a code repository.
Open Datasets Yes All models were trained using 64x64 images patches randomly chosen from a subset of the training set of Image Net (Deng et al. (2009)).
Dataset Splits No While the paper mentions the use of a 'small validation set', it does not provide specific details on the dataset split, such as exact percentages or sample counts for training, validation, and testing sets. It specifies using '12.5M patches' for training and 'images extracted from the test set of Image Net' for testing, but the validation split remains unquantified.
Hardware Specification Yes Reported times are average over fifty runs and were run on a Nvidia GTX Titan Black GPU, using 100 iterations of gradient decent.
Software Dependencies No The paper mentions using 'Adam (Kingma & Ba (2014))' for optimization but does not provide specific version numbers for Adam or any other software libraries or dependencies used.
Experiment Setup Yes As a baseline we use a CNN inspired in the one used by Dong et al. (2014). Specifically, a 4-layer CNN with {64, 64, 64, 32} feature maps and a linear output layer. Filter sizes are 7x7, 3x3, 3x3, and 5x5 respectively, with ReLU non-linearities at each hidden layer. The scattering network Ψ is a 3-layer complex convolutional network that uses complex modulus as point-wise non-linearities. Its filters are near-analytic Morlet wavelets that span 8 orientations and 3 scales. ... We used an architecture with { 32, 64, 64, 64, 219} feature maps with filter sizes {9x9, 9x9, 9x9, 3x3, 1x1} respectively. We alternate between optimizing Φ and Ψ. We adjust the learning rate corresponding to the parameters of Ψ by a factor η = 10^-4. We adjust the shrinkage of the total variation features at test time by adding a term λ r TV, with λ = 10^-8.