Projected variable three-term conjugate gradient algorithm for enhancing generalization performance in deep neural network training

doi:10.1016/j.neucom.2025.131568

Projected variable three-term conjugate gradient algorithm for enhancing generalization performance in deep neural network training

2025-09-14

0

PRE

AI

DOI：10.1016/j.neucom.2025.131568

Original

Original request for help

Save

Abstract

En 中文

• PVTTCG algorithm bridges convergence-generalization trade-off via orthogonal projection. • Achieves 0.33–3.92% gain on CIFAR and 35.9% loss drop in engineering task. • Demonstrates robust scalability with batch sizes up to 2,048 across applications. • Reveals systematic relationship between batch size and optimization performance. • Validates effectiveness from language modeling to 3D engineering predictions.

Projected variable three-term conjugate gradient algorithm for enhancing generalization performance in deep neural network training

Abstract

Journal

Organization