Convergence and Trade-Offs in Riemannian Gradient Descent and Riemannian Proximal Point

Authors: David Martı́nez-Rubio, Christophe Roux, Sebastian Pokutta

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

Reproducibility Variable Result LLM Response
Research Type Experimental Further, we explore beyond our theory with empirical tests. ... 5. Experiments: Numerical tests exploring beyond our theory. We observe that RGD presents a monotonic decrease in distance to an optimizer and show that RIPPA is competitive.
Researcher Affiliation Academia 1Zuse Institute Berlin, Germany 2Technische Universit at Berlin, Germany.
Pseudocode Yes Algorithm 1 Riemannian Inexact Proximal Point Algorithm (RIPPA)
Open Source Code No The paper mentions using the Pymanopt library (Townsend et al., 2016), published under the BSD-3-Clause license. This indicates the use of an open-source library, but the paper does not state that the authors are releasing their own code for the methodology described.
Open Datasets No The paper describes generating synthetic data ('centers yi') for the Karcher mean problem, but does not refer to a publicly available or open dataset with concrete access information (link, DOI, repository, or formal citation).
Dataset Splits No The paper describes a problem setup for the Karcher mean but does not provide specific details on training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions using the 'Pymanopt library (Townsend et al., 2016)' but does not specify its version number or any other software dependencies with specific version numbers.
Experiment Setup Yes We implement RGD with step sizes η = 1/L and η = 1/(LζR) as well as RIPPA performing a constant number of iterations of PRGD to approximately solve the proximal problems. ... We performed 3 iterations in each subroutine. ... We run until a fixed precision is reached in function value, and because of this, different algorithms stop at points at different distances from x*. ... In fact, we ran the algorithms for different settings, different initializations, and we performed a grid search on the step sizes.