Projection-free Distributed Online Learning with Sublinear Communication Complexity

Authors: Yuanyu Wan, Guanghui Wang, Wei-Wei Tu, Lijun Zhang

JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we perform simulation experiments on the multiclass classification problem and the binary classification problem to verify the performance of our proposed algorithms.
Researcher Affiliation Collaboration Yuanyu Wan EMAIL School of Software Technology, Zhejiang University Ningbo 315048, China Guanghui Wang EMAIL College of Computing, Georgia Tech Atlanta, GA 30332, USA Wei-Wei Tu EMAIL 4Paradigm Inc. Beijing 100000, China Lijun Zhang EMAIL National Key Laboratory for Novel Software Technology, Nanjing University Nanjing 210023, China
Pseudocode Yes The detailed procedures of our algorithm are presented in Algorithm 2, and it is called distributed block online conditional gradient (D-BOCG).
Open Source Code No The paper does not provide any explicit statement or link for open-source code for the methodology described.
Open Datasets Yes We conduct experiments on four publicly available data sets aloi, news20, a9a, and ijcnn1 from the LIBSVM repository (Chang and Lin, 2011), and the details of these data sets are summarized in Table 1.
Dataset Splits Yes For any data set, let Te denote the number of examples. We first divide it into n equally-sized parts where each part contains Te/n examples, and then distribute them onto n computing nodes in the network, where n = 9 for the multiclass classification problem and n = 100 for the binary classification problem.
Hardware Specification No The paper describes using 'n computing nodes' but does not provide any specific hardware details such as CPU/GPU models, memory, or cluster specifications.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers for its implementation.
Experiment Setup Yes For both methods, we simply initialize Xi(1) = 0v k, i [n]. According to Zhang et al. (2017), we set st = 1/t and η = c T 3/4 for D-OCG by tuning the constant c. Specifically, we set α = 0, K = L = T , and h = T 3/4/c by tuning the constant c. For both methods, the constant c is selected from {0.01, . . . , 1e5}.