Deep Network Classification by Scattering and Homotopy Dictionary Learning

Authors: John Zarka, Louis Thiry, Tomas Angles, Stephane Mallat

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

Reproducibility Variable Result LLM Response
Research Type Experimental Classification results are analyzed on Image Net. Section 4 analyzes image classification results of sparse scattering networks on Image Net 2012. Image Net 2012 (Russakovsky et al., 2015) is a challenging color image dataset of 1.2 million training images and 50,000 validation images, divided into 1000 classes. Table 1: Top 1 and Top 5 accuracy on Image Net with a same MLP classifier applied to different representations... Figure 4 left shows numerically that the ISTC algorithm for W = D minimizes the Lagrangian L(α) = 1/2 Dα β 2 + λ α 1 over α 0, with an exponential convergence which is faster than ISTA and FISTA. This is tested with a dictionary learned by minimizing the classification loss over Image Net.
Researcher Affiliation Collaboration John Zarka, Louis Thiry, Tomás Angles Département d informatique de l ENS, ENS, CNRS, PSL University, Paris, France {john.zarka,louis.thiry,tomas.angles}@ens.fr Stéphane Mallat Collège de France, Paris, France Flatiron Institute, New York, USA
Pseudocode No The paper includes architectural diagrams and mathematical equations but does not present structured pseudocode or algorithm blocks.
Open Source Code Yes Code to reproduce experiments is available at https://github.com/j-zarka/Sparse Scat Net
Open Datasets Yes Image Net 2012 (Russakovsky et al., 2015) is a challenging color image dataset of 1.2 million training images and 50,000 validation images, divided into 1000 classes.
Dataset Splits Yes Image Net 2012 (Russakovsky et al., 2015) is a challenging color image dataset of 1.2 million training images and 50,000 validation images, divided into 1000 classes.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU or CPU models, memory, or cloud computing instance types.
Software Dependencies No The paper mentions 'Py Torch implementation' and 'Kymatio' software package, but it does not specify version numbers for these software dependencies, which are necessary for reproducibility.
Experiment Setup Yes The ISTC network illustrated in Figure 2 has N = 12 layers with Re LU and no batch normalization. The sparse code is first calculated with a 1 1 convolutional dictionary D having 2048 vectors. The optimization learns a relatively large factor λ which yields a large approximation error LSx Dα1 / LSx 0.5, and a very sparse code α1 with about 4% non-zero coefficients. It is done with a stochastic gradient descent during 160 epochs using an initial learning rate of 0.01 with a decay of 0.1 at epochs 60 and 120.