Accelerated Convergence of Stochastic Heavy Ball Method under Anisotropic Gradient Noise

Authors: Rui Pan, Yuxing Liu, Xiaoyu Wang, Tong Zhang

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we fill this theoretical gap by establishing a non-asymptotic convergence bound for stochastic heavy-ball methods with step decay scheduler on quadratic objectives, under the anisotropic gradient noise condition.
Researcher Affiliation Academia 1The Hong Kong University of Science and Technology 2Fudan University 3University of Illinois Urbana-Champaign
Pseudocode Yes Algorithm 1 Multistage Stochastic Heavy Ball with minibatch
Open Source Code No The paper does not explicitly provide a link to its source code or state that it has been made open source for the methodology described.
Open Datasets Yes We use a4a1 dataset (Chang and Lin, 2011; Dua and Graff, 2017) to realize this setting... In this experiment, CIFAR-10 (Krizhevsky et al., 2009) dataset is adopted...
Dataset Splits Yes We use 5,000 randomly chosen samples in the training set to form a validation set, then conduct grid searches by training on the remaining 45,000 samples and selecting the hyperparameter with the best validation accuracy.
Hardware Specification No The paper mentions simulating distributed learning with '16 nodes' but does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers, such as library or solver names, needed to replicate the experiment.
Experiment Setup Yes In all of our experiments, we set the number of epochs to 100... we set different batch sizes M {2048, 512, 128}... For all schedulers, we set η0 {100, 10 1, 10 2, 10 3}. As for the choice of momentum factor β, we set β = 0.9 for stochastic heavy ball methods.