Sketching for Convex and Nonconvex Regularized Least Squares with Sharp Guarantees

Authors: Yingzhen Yang, Ping Li

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results validate the efficiency and effectiveness of both the SRO and Iterative SRO algorithms.
Researcher Affiliation Collaboration Yingzhen Yang School of Computing and Augmented Intelligence Arizona State University, Tempe, AZ 85281, USA EMAIL Ping Li Vec ML Inc., Bellevue, WA 98004, USA EMAIL
Pseudocode Yes Algorithm 1 Iterative SRO
Open Source Code No The paper describes algorithms (SRO, Iterative SRO) and their effectiveness but does not contain any explicit statement about making the source code publicly available or provide a link to a repository.
Open Datasets Yes Figure 6 and Figure 7 illustrate the accuracy (left) and NMI (right) of sketched Noisy SSC by SRO with respect to various choices of the regularization weight λ on the Extended Yale-B Dataset.
Dataset Splits No The paper describes generating synthetic data for some experiments (e.g., in Section C.1 and C.3) and mentions using the Extended Yale-B Dataset in Section C.4. However, it does not provide specific training/test/validation splits for any of the datasets used or generated. For the Extended Yale-B Dataset, it only mentions 'X is of size 1024 x 2414' without details on how it was split for experiments.
Hardware Specification Yes the running time is reported for γ = 3 on a CPU of Intel i5-11300H.
Software Dependencies No The paper mentions using Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) and Proximal Gradient Descent (PGD) but does not provide specific version numbers for these or any other software libraries, frameworks, or programming languages used for implementation.
Experiment Setup Yes Let M be the maximum number of iterations for FISTA, and we set M = 10000 for SRO and set M = 2000 for Iterative SRO... the maximum iteration number N for Iterative SRO in Algorithm 1 is always not greater than 5. (Section 7.1) We set λ = p log d/n. (Section C.1) We set λ = 0.1 p s log d/n, sparsity s = 3 log d. (Section C.3) We set α = 10λ, the sketch size n = n/5. (Section C.5)