Globally Q-linear Gauss-Newton Method for Overparameterized Non-convex Matrix Sensing

Authors: Xixi Jia, Fangchen FENG, Deyu Meng, Defeng Sun

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct experiments to demonstrate the effectiveness of the proposed AGN method for solving the over-parameterized non-convex matrix sensing problem. We set the ground truth matrix M = U ΣV , with U Rn r, V Rn r random orthogonal matrices and Σ is a diagonal matrix with condition number κ. We set n = 100, r = 5, d = 3r and the number of sensing matrices m = 50nr. All experiments were conducted using MATLAB on a Mac Book Pro with a 2.4 GHz quadcore Intel Core i5 CPU and 8 GB of memory.
Researcher Affiliation Academia Xixi Jia1 Fangchen Feng2 Deyu Meng3,4 Defeng Sun5 1School of Mathematics and Statistics, Xidian University 2L2TI Laboratory, University Sorbonne Paris Nord 3 School of Mathematics and Statistics, Xi an Jiaotong University 4 Macao Institute of Systems Engineering, Macau University of Science and Technology 5 Department of Applied Mathematics, The Hong Kong Polytechnic University
Pseudocode Yes A.1.2 The main AGN algorithm Algorithm 1: AGN for matrix sensing Data: A( ), b, η and the random Gauss initialization X0, t = 0. Result: The estimated solution X and the low-rank matrix ˆ M = PXX Q. while not end do t = t + 1; Update the approximated Gauss-Newton direction by Eq. (10); Update Xt and Xt+ 1 2 by Eq. (11); end
Open Source Code Yes The code for this paper is available at https://github.com/hsijiaxidian/AGN.
Open Datasets No We set the ground truth matrix M = U ΣV , with U Rn r, V Rn r random orthogonal matrices and Σ is a diagonal matrix with condition number κ. We set n = 100, r = 5, d = 3r and the number of sensing matrices m = 50nr.
Dataset Splits No The paper describes the generation of synthetic data and plots 'training curves' but does not specify explicit training, validation, or test dataset splits.
Hardware Specification Yes All experiments were conducted using MATLAB on a Mac Book Pro with a 2.4 GHz quadcore Intel Core i5 CPU and 8 GB of memory.
Software Dependencies No All experiments were conducted using MATLAB... (no specific version of MATLAB or other software libraries with version numbers are mentioned).
Experiment Setup Yes We set the ground truth matrix M = U ΣV , with U Rn r, V Rn r random orthogonal matrices and Σ is a diagonal matrix with condition number κ. We set n = 100, r = 5, d = 3r and the number of sensing matrices m = 50nr. All the competing methods are initialized with random Gaussian matrix with zero mean the variance 1/n.