Federated Sampling with Langevin Algorithm under Isoperimetry

Authors: Lukang Sun, Adil Salim, Peter Richtárik

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental A Experiments We now experiment our algorithm and compare it to QLSD++ (Vono et al., 2022). Since this paper is theoretical, we chose a simple one dimensional example, in order to be able to visualize the performance of the algorithms on histograms.
Researcher Affiliation Collaboration Lukang Sun EMAIL King Abdullah University of Science and Technology Adil Salim EMAIL Microsoft Research Peter Richtárik EMAIL King Abdullah University of Science and Technology
Pseudocode Yes Algorithm 1 MARINA (Gorbunov et al., 2021) ... Algorithm 2 Langevin-MARINA (proposed algorithm)
Open Source Code No The paper does not provide any explicit statement about the availability of source code or a link to a code repository.
Open Datasets No The target distribution is π exp( F), where F is defined by a mathematical function... This is a synthetic setup, not using an external dataset.
Dataset Splits No The paper uses a synthetic target distribution for its experiments, therefore, traditional dataset splits are not applicable.
Hardware Specification No The paper describes experimental setup parameters like the number of clients and steps, but it does not specify any hardware details such as CPU/GPU models or memory.
Software Dependencies No The paper mentions several algorithms and components like MARINA, QLSD++, and a compression operator, but it does not specify any software versions for these or other dependencies.
Experiment Setup Yes Experimental setup. We use n = 5 clients, and client i has N(i) subfunctions Fij. The number N(i) is randomly chosen between 10 and 20. ... In Langevin-Marina we set p = 0.001, i.e., a full gradient is computing every 1000 iterations in expectation. In QLSD++ (Vono et al., 2022) we set α = 0.2, the initial memory term η(i) = 0, i [5] and l = 1000, i.e., a full gradient is computing every 1000 iterations. We use the compression operator from Vono et al. (2022) with quantization level s = 28. ... We run 400000 steps of both algorithms and collect all the points generated by each algorithm to plot their histogram.