A Pseudo-Euclidean Iteration for Optimal Recovery in Noisy ICA

Authors: James R. Voss, Mikhail Belkin, Luis Rademacher

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental Results We compare the proposed PEGI algorithm with existing ICA algorithms. In addition to qorth+GI-ICA (i.e., GI-ICA with quasi-orthogonalization for preprocessing), we use the following baselines: JADE [3] is a popular fourth cumulant based ICA algorithm designed for the noise free setting. We use the implementation of Cardoso and Souloumiac [5].
Researcher Affiliation Academia James Voss The Ohio State University vossj@cse.ohio-state.edu Mikhail Belkin The Ohio State University mbelkin@cse.ohio-state.edu Luis Rademacher The Ohio State University lrademac@cse.ohio-state.edu
Pseudocode Yes Algorithm 1 Recovers a column of A up to a scaling factor if u0 is generically chosen. Inputs: Unit vector u0, C, f k 1 repeat uk f(C uk 1)/ f(C uk 1) k k + 1 until Convergence (up to sign) return uk
Open Source Code No The paper provides links to implementations of baseline algorithms (JADE and Fast ICA) used for comparison, e.g., 'JADE [3] is a popular fourth cumulant based ICA algorithm designed for the noise free setting. We use the implementation of Cardoso and Souloumiac [5]. http://perso.telecom-paristech.fr/cardoso/Algo/Jade/jadeR.m, 2005.' and 'Matlab Fast ICA v 2.5. http://research.ics.aalto.fi/ica/fastica/code/dlcode.shtml, 2005.' However, it does not state that the code for the proposed PEGI algorithm is open-source or provide a link to it.
Open Datasets No The paper uses simulated data, describing how it was generated ('We constructed mixing matrices A with condition number 3 via a reverse singular value decomposition... We drew data from a noisy ICA model X = AS + η... We generated 100 matrices A for our experiments with 100 corresponding ICA data sets for each sample size and noise power.'). It does not provide access information (link, DOI, citation) for a publicly available dataset.
Dataset Splits No The paper mentions varying 'sample size' for experiments, but does not specify explicit training, validation, or test dataset splits, nor does it refer to predefined splits or cross-validation.
Hardware Specification No No specific hardware details (such as GPU/CPU models or memory specifications) used for running the experiments were provided in the paper.
Software Dependencies Yes The paper states: 'We use the implementation of Cardoso and Souloumiac [5]. Matlab JADE for real-valued data v 1.8. http://perso.telecom-paristech.fr/ cardoso/Algo/Jade/jadeR.m, 2005.' and 'We use the implementation of G avert et al. [10]. Matlab Fast ICA v 2.5. http://research.ics. aalto.fi/ica/fastica/code/dlcode.shtml, 2005.' These provide specific version numbers for the software used for baseline comparisons.
Experiment Setup Yes We constructed mixing matrices A with condition number 3 via a reverse singular value decomposition (A = UΛV T ). The matrices U and V were random orthogonal matrices, and Λ was chosen to have 1 as its minimum and 3 as its maximum singular values, with the intermediate singular values chosen uniformly at random. We drew data from a noisy ICA model X = AS + η where cov(η) = Σ was chosen to be malaligned with cov(AS) = AAT . We set Σ = p(10I AAT ) where p is a constant defining the noise power. ... The source distributions used in our ICA experiments were the Laplace and Bernoulli distribution with parameters 0.05 and 0.5 respectively, the t-distribution with 3 and 5 degrees of freedom respectively, the exponential distribution, and the uniform distribution. Each distribution was normalized to have unit variance, and the distributions were each used twice to create 14-dimensional data.