Recovering Simultaneously Structured Data via Non-Convex Iteratively Reweighted Least Squares

Authors: Christian Kümmerle, Johannes Maly

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The numerical simulations in Section 4 support our theoretical findings and provide empirical evidence for the efficacy of the proposed method.
Researcher Affiliation Academia Christian Kümmerle Department of Computer Science University of North Carolina at Charlotte Charlotte, NC 28223, USA kuemmerle@uncc.edu Johannes Maly Department of Mathematics Ludwig-Maximilians-Universität München 80799 Munich, Germany and Munich Center for Machine Learning (MCML) maly@math.lmu.de
Pseudocode Yes Algorithm 1 IRLS for simultaneously low-rank rand row-sparse matrices
Open Source Code Yes We refer to the MATLAB implementation available in the repository https://github.com/ckuemmerle/simirls for further details.
Open Datasets No In the experiments, we chose random ground truths X Rn1 n2 of rank r and row-sparsity s such that X = X / X F , where X = U diag(d )V , and where U Rn1 r is a matrix with s non-zero rows whose location is chosen uniformly at random and whose entries are drawn from i.i.d. standard Gaussian random variables, d has i.i.d. standard Gaussian entries and V Rn2 r has likewise i.i.d. standard Gaussian entries.
Dataset Splits No The paper describes experiments like 'phase transition experiments' and averaging over '64 random trials' but does not specify train/validation/test dataset splits.
Hardware Specification Yes The CPU models used in the simulations are Dual 18-Core Intel Xeon Gold 6154, Dual 24-Core Intel Xeon Gold 6248R, Dual 8-Core Intel Xeon E5-2667, 28-Core Intel Xeon E5-2690 v3, 64-Core Intel Xeon Phi KNL 7210-F.
Software Dependencies Yes The experiments of Section 4 were conducted using MATLAB implementations of the three algorithms on different Linux machines using MATLAB versions R2019b or R2022b.
Experiment Setup Yes In all phase transition experiments, we define successful recovery such that the relative Frobenius error X(K) X F / X F of the iterate X(K) returned by the algorithm relative to the simultaneously low-rank and row-sparse ground truth matrix X is smaller than the threshold 10 4. As stopping criteria, we used the criterion that the relative change of Frobenius norm satisfies X(k) X(k 1) F / X(k) F < tol for IRLS, the change in the matrix factors norms satisfy Uk Uk 1 < tol and Vk Vk 1 < tol for SPF, and the norm of the Riemannian gradient in Riem Ada IHT being smaller than tol for tol = 10 10, or if a maximal number of iterations is reached. This iteration threshold was chosen as max_iter = 250 for IRLS and SPF and as max_iter = 2000 for Riem Ada IHT.