Work

Unrolling Nonnegative Matrix Factorization via Learning Single Optimization Step in Alternating Optimization

Year: 2025

Type: conference-paper

Abstract: Unrolling, a technique within the learning-to-optimize framework, seeks to enhance optimization speed and effectiveness through supervised learning. While conventional unrolling methods have pridominantly focused on convex optimization problems, their application to nonconvex problems remains limited. This paper proposes an unrolling method for nonnegative matrix factorization (NMF), a simple linear model that requires alternating optimization to decompose a nonnegative data matrix into the product of two nonnegative matrices. Our method, termed Neural Adaptive Gradient NMF (NAG-NMF), employs neural networks to estimate the update vectors at each optimization step. We also introduce an efficient training strategy for NAG-NMF and demonstrate its effectiveness through experimental evaluations.

Source: International Conference on Intelligent Systems, Metaheuristics & Swarm Intelligence (ISMSI)

DOI: 10.1145/3760622.3760628

Author: Toshimitsu Aritake

Institution: Hitotsubashi University

Language: English

Topic: Stochastic Gradient Optimization Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funder: Japan Society for the Promotion of Science

Award: Development of a Fast and Accurate Computational Method through Learning-based Iterative Alternating Optimization