Inertial Block Proximal Methods for Non-Convex Non-Smooth Optimization

Authors: Hien Le, Nicolas Gillis, Panagiotis Patrinos

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

Reproducibility Variable Result LLM Response
Research Type Experimental We deploy the proposed methods to solve non-negative matrix factorization (NMF) and show that they compete favourably with the state-of-the-art NMF algorithms. Additional experiments on non-negative approximate canonical polyadic decomposition, also known as non-negative tensor factorization, are also provided.
Researcher Affiliation Academia 1Department of Mathematics and Operational Research, University of Mons, Belgium 2Department of Electrical Engineering (ESAT-STADIUS), KU Leuven, Belgium.
Pseudocode Yes Algorithm 1 IBP and IBPG
Open Source Code Yes The code is available at https://github.com/Le Thi Khanh Hien/IBPG
Open Datasets Yes We test the algorithms on two widely used hyperspectral images, namely the Urban and San Diego data sets; see (Gillis et al., 2015).
Dataset Splits No The paper mentions generating synthetic data and using random initializations but does not specify explicit train/validation/test dataset splits with percentages or counts.
Hardware Specification Yes All tests are preformed using Matlab R2015a on a laptop Intel CORE i78550U CPU @1.8GHz 16GB RAM.
Software Dependencies Yes All tests are preformed using Matlab R2015a on a laptop Intel CORE i78550U CPU @1.8GHz 16GB RAM.
Experiment Setup Yes 3.3. Choice of Parameters for NMF Let us illustrate the choice of parameters for NMF. In the remainder of this paper, in the context of NMF, we will refer to IBPG as Algoritm 1 with the choice Tk = 2 (cyclic update of U and V ), and to IBPG-A with the choice Tk > 2 (U and V are updated several times). For IBPG and IBPG-A, we take L(k,m) 1 = L(k) 1 = ( V (k 1))T V (k 1) and L(k,m) 2 = L(k) 2 = ( U (k))T U (k) for m 1. We take β(k,m) i = 1/ L(k) i , γ(k,m) i = min τk 1 L(k 1) i L(k) i and α(k,m) i = α γ(k,m) i , where τ0 = 1, τk = 1 2(1 + q 1 + 4τ 2 k 1), γ = 0.99 and α = 1.01. Regarding IBP, we choose 1/β(k,m) i = 0.001 and α(k,m) i = α(k) = min( β, γ α(k 1)), with β = 1, γ = 1.01 and α(1) = 0.6.