Projection-Free Methods for Solving Nonconvex-Concave Saddle Point Problems

Authors: Morteza Boroun, Erfan Yazdandoost Hamedani, Afrooz Jalilzadeh

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we implement our methods to solve Robust Multiclass Classification problem described in Example 1 and Dictionary Learning problem in Example 2. As shown in Figure 1, our algorithms outperform the competing approaches
Researcher Affiliation Academia Morteza Boroun , Erfan Yazdandoost Hamedani , Afrooz Jalilzadeh Department of Systems & Industrial Engineering The University of Arizona Tucson, AZ 85721 morteza@arizona.edu, erfany@arizona.edu, afrooz@arizona.edu
Pseudocode Yes Algorithm 1 Regularized Primal-dual Conditional Gradient (R-PDCG) method Algorithm 2 Conditional Gradient with Regularized Projected Gradient Ascent (CG-RPGA)
Open Source Code No The paper does not provide concrete access to source code for the methodology described, nor does it state that the code is released or available in supplementary materials.
Open Datasets Yes We conduct experiments on rcv1 dataset (n = 15564, d = 47236, k = 53) and news20 dataset (n = 15935, d = 62061, k = 20) from LIBSVM repository1. 1https://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets
Dataset Splits No The paper mentions the datasets used (rcv1 and news20 from LIBSVM repository) but does not provide specific details on how the data was split into training, validation, and test sets.
Hardware Specification No The paper does not provide specific hardware details (like GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions using datasets from the LIBSVM repository but does not provide specific ancillary software details with version numbers needed to replicate the experiment.
Experiment Setup Yes For all the algorithms, the step-sizes are selected as suggested by their theoretical result and scaled to have the best performance. In particular, for R-PDCG we let τ = 10 K5/6 and µ = 10 3 K1/6 ; for CG-RPGA we let τ = 10 K3/4 and K1/4 ; for AGP we let the primal step-size 1 k, dual step-size as 0.2, and the dual regularization parameter as 10 1 k1/4 ; for SPFW both primal and dual step-sizes are selected to be diminishing as 2 k+2.