A Novel M-Estimator for Robust PCA

Authors: Teng Zhang, Gilad Lerman

JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our method with many other algorithms for robust PCA on synthetic and real data sets and demonstrate state-of-the-art speed and accuracy. Keywords: principal components analysis, robust statistics, M-estimator, iteratively re-weighted least squares, convex relaxation (...) 6. Numerical Experiments
Researcher Affiliation Academia Teng Zhang EMAIL Institute for Mathematics and its Applications University of Minnesota Minneapolis, MN 55455, USA Gilad Lerman EMAIL School of Mathematics University of Minnesota Minneapolis, MN 55455, USA
Pseudocode Yes Algorithm 1 The Geometric Median Subspace Algorithm (...) Algorithm 2 Practical and Regularized Minimization of (4) (...) Algorithm 3 The Extended Geometric Median Subspace Algorithm
Open Source Code Yes Supp. webpage: http://www.math.umn.edu/~lerman/gms.
Open Datasets Yes We used the images of the first two persons in the extended Yale face database B (Lee et al., 2005) (...) We consider the following two videos used by Cand es et al. (2011): Lobby in an office building with switching on / offlights and Shopping center from http://perception.i2r.a-star.edu.sg/bk_ model/bk_index.html.
Dataset Splits No The paper describes generating synthetic data with N1 inliers and N0 outliers, and using full real-world datasets (Yale Face database, video frames) for evaluation, but does not specify explicit training/test/validation splits or cross-validation setups for reproducing experiments.
Hardware Specification Yes We ran the experiments on a computer with Intel Core 2 CPU at 2.66GHz and 2 GB memory. (...) We used a computer with Intel Core 2 Quad Q6600 2.4GHz and 8 GB memory due to the large size of these data.
Software Dependencies No The paper lists specific codes and implementations of various algorithms (OP, HR-PCA, MKF, PCP, MDR, LLD) and states their own code will appear on a supplemental webpage, but it does not provide specific version numbers for programming languages, libraries, or other software dependencies.
Experiment Setup Yes We fix the regularization parameter to be smaller than the rounding error, that is, δ = 10 20 (...) For LLD, OP and PCP we set the mixture parameter λ as 0.8 p max(D, N) respectively (following the suggestions of Mc Coy and Tropp (2011) for LLD/OP and Cand es et al. (2011) for PCP). These choices of parameters are also used in experiments with real data sets in 6.7 and 6.8. For the common M-estimator, we used u(x) = 2 max(ln(x)/x, 1030) and the algorithm discussed by Kent and Tyler (1991).