Asynchronous Stochastic Quasi-Newton MCMC for Non-Convex Optimization

Authors: Umut Simsekli, Cagatay Yildiz, Than Huy Nguyen, Taylan Cemgil, Gael Richard

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform several experiments on both synthetic and real datasets. The results support our theory and show that the proposed algorithm provides a significant speedup over the recently proposed synchronous distributed L-BFGS algorithm.
Researcher Affiliation Academia 1LTCI, T el ecom Paris Tech, Universit e Paris-Saclay, 75013, Paris, France 2Department of Computer Science, Aalto University, Espoo, 02150, Finland 3Department of Computer Engineering, Bo gazic i University, 34342, Bebek, Istanbul, Turkey.
Pseudocode Yes Algorithm 1: as-L-BFGS: Master node; Algorithm 2: as-L-BFGS: Worker node (w)
Open Source Code No The paper mentions 'This code can be used both in a distributed environment or a single computer with multiprocessors' but does not provide an explicit statement of public release or a link to the source code for the described methodology.
Open Datasets Yes Movie Lens 1Million (ML-1M), 10Million (ML-10M), and 20Million (ML20M) (grouplens.org)
Dataset Splits No The paper does not explicitly provide specific training/test/validation dataset splits, percentages, or sample counts.
Hardware Specification No We have conducted these experiments on a cluster of more than 500 interconnected computers, each of which is equipped with variable quality CPUs and memories.
Software Dependencies No For real data experiments, we have implemented all the three algorithms in C++ by using a low-level message passing protocol for parallelism, namely the Open MPI library. This does not specify version numbers for C++ or Open MPI.
Experiment Setup Yes In these experiments, we set K = 5 for all the three datasets and we set the number of workers to W = 10.