Discrete Deep Feature Extraction: A Theory and New Architectures

Authors: Thomas Wiatowski, Michael Tschannen, Aleksandar Stanic, Philipp Grohs, Helmut Boelcskei

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on handwritten digit classification and facial landmark detection including feature importance evaluation complement the theoretical findings.
Researcher Affiliation Academia Thomas Wiatowski1 WITHOMAS@NARI.EE.ETHZ.CH Michael Tschannen1 MICHAELT@NARI.EE.ETHZ.CH Aleksandar Stani c1 ASTANIC@STUDENT.ETHZ.CH Philipp Grohs2 PHILIPP.GROHS@UNIVIE.AC.AT Helmut B olcskei1 BOELCSKEI@NARI.EE.ETHZ.CH 1Dept. IT & EE, ETH Zurich, Switzerland 2Dept. Math., University of Vienna, Austria
Pseudocode No The paper focuses on mathematical theory and experimental results but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code available at http://www.nari.ee.ethz.ch/ commth/research/
Open Datasets Yes We use the MNIST dataset of handwritten digits (Le Cun & Cortes, 1998) which comprises 60,000 training and 10,000 test images of size 28x28. ... We use the Caltech 10,000 Web Faces data base (Angelova et al., 2005).
Dataset Splits Yes The MNIST dataset of handwritten digits (Le Cun & Cortes, 1998) which comprises 60,000 training and 10,000 test images... The penalty parameter of the SVM and the localization parameter of the RBF kernel are selected via 10-fold cross-validation for each combination of wavelet filter, non-linearity, and pooling operator. ... We select 80% of the images uniformly at random to form a training set and use the remaining images for testing.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments.
Software Dependencies No The paper mentions software components like 'SVM', 'random forests (RF)', 'algorithme a trous', and 'Viola-Jones face detector', but does not provide specific version numbers for any of these.
Experiment Setup Yes We set D = 3, and compare different network configurations, each defined by a single module (i.e., we use the same filters, non-linearity, and pooling operator in all layers). Specifically, we consider Haar wavelets and reverse biorthogonal 2.2 (RBIO2.2) wavelets (Mallat, 2009), both with J = 3 scales, the non-linearities described in Section 2.2.1, and the pooling operators described in Section 2.3.1 (with S1 = 1 and S2 = 2). We use a SVM with radial basis function (RBF) kernel for classification. ... In both cases, we fix the number of trees to 30 and select the tree depth using out-of-bag error estimates... For the feature extractor ΦΩ we set D = 4, employ Haar wavelets with J = 3 scales and the modulus non-linearity in every network layer, no pooling in the first layer and average pooling with uniform weights 1/S2 d, Sd = 2, in layers d = 2, 3.