Researchers have developed a new Adaptive Hybrid Subspace Levenberg-Marquardt (HSLM) algorithm designed to efficiently solve large-scale nonlinear least-squares problems. This method constructs a low-dimensional subspace using various information sources, including gradient and curvature data, to compute a damped LM step. A key innovation is an adequacy monitor that adaptively enriches the subspace when needed, and a decoupled step acceptance strategy using Armijo backtracking. Numerical experiments on neural network training problems demonstrate that HSLM achieves convergence comparable to existing methods while significantly reducing computational costs, especially for problems with a high parameter dimension. AI
IMPACT Potentially accelerates training for large neural networks by improving optimization efficiency.
RANK_REASON Academic paper detailing a new numerical algorithm. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
