Projection techniques to update the truncated SVD of evolving matrices with applications

Authors: Vasileios Kalantzis, Georgios Kollias, Shashanka Ubaru, Athanasios N. Nikolakopoulos, Lior Horesh, Kenneth Clarkson

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we benchmark the performance of Algorithm 1 on problems from the areas of LSI and face recognition. The experiments were conducted in a Matlab environment (version R2020a), using 64-bit arithmetic, on a single core of a computing system equipped with an 2.5GHz Quad-Core Intel Core i7 processor and 16 GB of system memory.
Researcher Affiliation Industry 1IBM Research, Thomas J. Watson Research Center, Yorktown Heights, NY 10598 2Amazon, 550 Terry Ave N, Seattle, WA 98109 (work done prior to joining Amazon) 3IBM Research, Almaden Research Center, San Jose, CA 95120.
Pseudocode Yes Algorithm 1 The proposed algorithm (row updates). and Algorithm 2 The proposed algorithm (column updates).
Open Source Code Yes Open source implementation of the method is available at: https://github.com/nikolakopoulos/Truncated SVDupdates.
Open Datasets Yes Table 2. Text mining test matrices (obtained by http://web.eecs.utk.edu/research/lsi/ and https://github.com/ZJULearning/MatlabFunc). Our experiments are performed on the Yale and AT&T face databases.3 These files were obtained in a Matlab-ready format from http://www.cad.zju.edu.cn/home/dengcai/Data/Face Data.html; see (Cai et al., 2007a;b)
Dataset Splits Yes The default values used throughout our experiments are µ = m/10 and φ = 10, i.e., the initial matrix B is formed by the leading 10% of the rows matrix A, and the remaining 90% of rows is added in ten batches, with each batch roughly representing 9% of the total rows of matrix A. Table 5 lists the classification mean error-rate averaged over fifty splits, where each split contains eight images per individual in the training set.
Hardware Specification Yes The experiments were conducted in a Matlab environment (version R2020a), using 64-bit arithmetic, on a single core of a computing system equipped with an 2.5GHz Quad-Core Intel Core i7 processor and 16 GB of system memory.
Software Dependencies Yes The experiments were conducted in a Matlab environment (version R2020a), using 64-bit arithmetic...
Experiment Setup Yes The default values used throughout our experiments are µ = m/10 and φ = 10, i.e., the initial matrix B is formed by the leading 10% of the rows matrix A, and the remaining 90% of rows is added in ten batches, with each batch roughly representing 9% of the total rows of matrix A. The number of computed eigenfaces is set as k = 25 and k = 50. For this task we use ρ-nearest neighbors with ρ = 5 (Fix, 1985).