Zeno++: Robust Fully Asynchronous SGD

Authors: Cong Xie, Sanmi Koyejo, Indranil Gupta

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Zeno++ outperforms existing Byzantine-tolerant asynchronous SGD algorithms. We conduct experiments on two benchmarks: CIFAR-10 image classification dataset (Krizhevsky, 2009), and Wiki Text-2 language modeling dataset (Merity et al., 2017). Our empirical results show good performance compared to previous work.
Researcher Affiliation Academia 1Department of Computer Science, University of Illinois, Urbana-Champaign, USA.
Pseudocode Yes Algorithm 1 Zeno+. Algorithm 2 Zeno++.
Open Source Code No The detailed network architecture can be found in our submitted source code (will be released upon publication).
Open Datasets Yes We conduct experiments on two benchmarks: CIFAR-10 image classification dataset (Krizhevsky, 2009), and Wiki Text-2 language modeling dataset (Merity et al., 2017).
Dataset Splits Yes From the training set, we randomly extracted 2.5k of them as the validation set for Zeno++, the remaining are randomly partitioned onto all the workers.
Hardware Specification No This work was funded in part by the following grants: NSF IIS 1909577, NSF CNS 1908888, and a JP Morgan Chase Fellowship, along with computational resources donated by Intel, AWS, and Microsoft Azure.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes In all the experiments, we take the learning rate γ = 0.1, mini-batch size n = ns = 128, ρ = 0.002, ϵ = 0.1, k = 10. In all the experiments, we take the learning rate γ = 20, mini-batch size n = ns = 20, k = kw = 10, ρ = 10, ϵ = 2.