Escaping from saddle points on Riemannian manifolds

Authors: Yue Sun, Nicolas Flammarion, Maryam Fazel

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our numerical experiment, we choose H to be a diagonal matrix H = diag(0, 1, 2, 3, 4) and let k = 3. The Euclidean basis (ei) are an eigenbasis of H and the first order stationary points of the objective function are [ei1, ei2, ei3]G with distinct basis and G being unitary. The local minimizers are [e3, e4, e5]G. We start the iteration at X0 = [e2, e3, e4] and see in Fig. 3 the algorithm converges to a local minimum.
Researcher Affiliation Academia Yue Sun University of Washington Seattle, WA 98105 yuesun@uw.edu Nicolas Flammarion EPFL Lausanne, Switzerland nicolas.flammarion@epfl.ch Maryam Fazel University of Washington Seattle, WA 98105 mfazel@uw.edu
Pseudocode Yes Algorithm 1 Perturbed Riemannian gradient algorithm
Open Source Code No The paper mentions 'Manopt, a Matlab toolbox for optimization on manifolds' and other 'open-source packages' but does not provide a specific link or statement about the open-sourcing of the code implemented for this paper's methodology.
Open Datasets No The paper describes experiments on a 'k PCA problem' with a constructed matrix H = diag(0, 1, 2, 3, 4) and a 'Burer-Monteiro approach' with a constructed sparse matrix A R100 20. It does not mention using or providing access to any publicly available or open datasets.
Dataset Splits No The paper does not provide specific dataset split information (e.g., percentages, sample counts, or citations to predefined splits) for training, validation, or testing.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions 'Manopt, a Matlab toolbox for optimization on manifolds' and 'geomstats: a python package' but does not provide specific version numbers for any ancillary software used in its experiments.
Experiment Setup Yes Set constants: ˆc 4, C := C(K, β, ρ) (defined in Lemma 2 and proof of Lemma 8) and cmax 1 56ˆc2 , r = cmax χ2 ϵ, χ = 3 max{log( dβ(f(x0) f ) ˆcϵ2δ ), 4}. Set threshold values: fthres = cmax ρ , gthres = cmax χ2 ϵ, tthres = χ cmax β ρϵ, tnoise = tthres 1. Set stepsize: η = cmax β .