Information Geometry and Minimum Description Length Networks

Authors: Ke Sun, Jun Wang, Alexandros Kalousis, Stephan Marchand-Maillet

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present a new learning machine with theories, algorithms, and simulations. Section 5 shows density estimation simulations. Figure 5. Average testing error over all instance datasets. Figure 6. Average training (resp. testing) errors over the training (resp. testing) samples on digit1 against n1 (n2 is fixed to 1) with a training : testing ratio of 1 : 9.
Researcher Affiliation Collaboration Ke Sun SUNK.EDU@GMAIL.COM Viper Group, Computer Vision and Multimedia Laboratory, University of Geneva, Switzerland Jun Wang JWANG1@EXPEDIA.COM Expedia, Switzerland Alexandros Kalousis ALEXANDROS.KALOUSIS@HESGE.CH Business Informatics Department, University of Applied Sciences, Western Switzerland Department of Computer Science, University of Geneva, Switzerland St ephane Marchand-Maillet STEPHANE.MARCHAND-MAILLET@UNIGE.CH Viper Group, Computer Vision and Multimedia Laboratory, University of Geneva, Switzerland
Pseudocode Yes Alg. 1: N, A = HARDN ({xi}n i=1, n1, . . . , n L, γ) Alg. 2: relocate η, {(ηL i , w L i )}, {(ηR i , w R i )}, γ
Open Source Code Yes The codes are at https://git.unige.ch/gitweb/ marchand/mdlnetworks
Open Datasets Yes The datasets used are listed in table 1. Name: faithful, 2moons, 9blobs, iris, wine, digit1. Both from the UCI repository. Hand-written digits of 1 in MNIST.
Dataset Splits Yes For each dataset, a large number of instance datasets are generated based on different random seeds and different splits of training and testing data. with a training : testing ratio of 1 : 9.
Hardware Specification No The paper does not provide any specific hardware details used for running the experiments.
Software Dependencies No The paper mentions 'scikit-learn machine learning library (Pedregosa et al., 2011)' but does not provide specific version numbers for this or any other software dependency.
Experiment Setup Yes In alg. 1, the µ s in Ll (1 l L) are initialized (line 1) by the k-means (Arthur & Vassilvitskii, 2007) centroids of the µ s in Ll 1; the Σ s are either initialized by the covariance of the k-means clusters, or the global data covariance. The learning rate γ is adjusted (line 11) online by a bold driver (Battiti, 1989).