Approximate Frank-Wolfe Algorithms over Graph-structured Support Sets

Authors: Baojian Zhou, Yifan Sun

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results suggest that even these improved bounds are pessimistic, showing fast convergence in recovering real-world images with graph-structured sparsity. We empirically demonstrate that the proposed methods are more effective and efficient compared with PGD-based and MP-based methods on the graph-structured linear regression problem. We empirically evaluate our methods through the task of graph-structured linear regression problem over several graph-structured images.
Researcher Affiliation Academia 1School of Data Science, Fudan University, Shanghai, China 2Department of Computer Science, Stony Brook University, Stony Brook, New York, USA.
Pseudocode Yes Algorithm 1 FW-type methods for GSCOs
Open Source Code Yes Our code and datasets will be made publically available upon publication, and is included in the submission. Our code and datasets are accessible at https://github.com/baojian/dmo-fw.
Open Datasets Yes Our code and datasets will be made publically available upon publication, and is included in the submission. We use two datasets: 1) 10 MNIST images. ... 2) Angio image. We also choose a sparse angio image from (Hegde et al., 2015b).
Dataset Splits No The paper specifies number of samples (n = 2.5 | supp(x )| samples, n = 5 | supp(x )| samples) and number of trials (20 trials) for experiments, but does not explicitly describe train/validation/test splits, percentages, or cross-validation methodology.
Hardware Specification No We run all methods on a sever with 246GB memory and 80 cores.
Software Dependencies Yes All methods are implemented in Python-3.8. The graph projection operator is implemented in C++11.
Experiment Setup Yes The step size of both DMO-FW and DMO-Acc FW are set to ηt = 2/(t + 2) for all t 0. The DMO we used is the head projection of Hegde et al. (2015b). We use Option I for DMO-ACCFW and simply set L = 1. Specifically, we pick n = 2.5 | supp(x )| samples. We run each experiment for 20 trials, and compare our methods against the generalized MP (GEN-MP) discussed in Locatello et al. (2018).