Improved Powered Stochastic Optimization Algorithms for Large-Scale Machine Learning

Authors: Zhuang Yang

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical tests using various benchmark data sets indicate the efficiency and robustness of our proposed algorithms.
Researcher Affiliation Academia Zhuang Yang EMAIL School of Computer Science and Technology Soochow University Suzhou, 215006, China
Pseudocode Yes Algorithm 1 PB-SVRGE Algorithm 2 PB-SVRGE-RSBB Algorithm 3 PB-SGD-RSBB
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes In order to evaluate the proposed algorithms, we conducted the experiments on six standard data sets, where the details of these data sets were listed in Table 1. 1 Specifically, the data sets (a8a, covtype, ijcnn1 and news20) can be downloaded from LIBSVM (Chang and Lin, 2011). In addition, we took the non-convex logistic regression as the loss function. ... 1. For the CIFAR-10 data set, it can be downloaded from http://www.cs.toronto.edu/~kriz/cifar.html. For the MNIST data set, it can be accessed from http://yann.lecun.com/exdb/mnist/.
Dataset Splits No The paper mentions using several datasets (a8a, covtype, CIFAR-10, ijcnn1, MNIST, news20.binary) but does not provide specific details on how these datasets were split into training, validation, or test sets. For example, it only lists the total number of examples for each dataset in Table 1.
Hardware Specification Yes Note that all experiments were conducted on an Intel(R) Core(TM) i7-10750H CPU @2.60GHz 2.59GHz with MATLAB 2019a.
Software Dependencies Yes Note that all experiments were conducted on an Intel(R) Core(TM) i7-10750H CPU @2.60GHz 2.59GHz with MATLAB 2019a.
Experiment Setup Yes We studied the numerical behaviors of the proposed algorithms with the regularizer coefficient λ = 10^-1. ... On all data sets, we set b = 10. In addition, on a8a and i jcnn1, we set η = 0.01. While on covtype and news20.binary, we set η = 0.1. ... we set b = 10, b H = 20, ζ = 1 and γ = 0.9 on different data sets when executing PB-SVRGE-RSBB (Algorithm 2). The parameter ζ was chosen from {0.01, 0.1, 1, 10}.