Online Convex Optimization in the Random Order Model

Authors: Dan Garber, Gal Korcia, Kfir Levy

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments We conduct two sets of experiments demonstrating our theoretical findings for the online PCA problem. (Recall also our experiment from the introduction, which considers the online linear regression problem.)
Researcher Affiliation Academia 1Department of Industrial Engineering and Management, Technion Israel Institute of Technology, Haifa, Israel 2Department of Electrical Engineering, Technion Israel Institute of Technology, Haifa, Israel.
Pseudocode Yes Algorithm 1 batch Online Principal Component Analysis based on Online Gradient Ascent
Open Source Code No The paper does not contain any explicit statement about providing open-source code for the described methodology, nor does it provide any links to a code repository.
Open Datasets Yes MNIST we use the training set of the MNIST handwritten digit recognition dataset, which contains 60,000 28 28 images, which we split into 58200 images for testing, and 1800 images (3%) are used to compute the initialization W0 and step sizes ηt = 1 103 t.
Dataset Splits No MNIST we use the training set of the MNIST handwritten digit recognition dataset, which contains 60,000 28 28 images, which we split into 58200 images for testing, and 1800 images (3%) are used to compute the initialization W0 and step sizes ηt = 1 103 t.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes The data is sampled from a multi-variate Normal distribution with zero mean and diagonal covariance matrix Σ. For each value of k, we have Σi,i = 1 for 1 i k and Σi,i = gap 2 i 0.1 for k+1 i d. In our experiments gap = 0.1, k = {1, 2, 3, 7}, d = 1000, and the window size is L = 10. We use 3% of the data to compute the initialization W0, and step sizes ηt = 1 t. ... We set L = 20 and k = {1, 3, 7, 15}.