Two-phase collaborative model compression training for joint pruning and quantization

doi:10.1016/j.neunet.2025.108506

Two-phase collaborative model compression training for joint pruning and quantization

2025-12-23

0

PRE

AI

DOI：10.1016/j.neunet.2025.108506

Original

Original request for help

Save

Abstract

En 中文

• We propose a unified collaborative model compression training to integrate pruning, quantization and performance objectives, balancing complexity reduction and precision loss in training. • A novel constraint function combining sparse regularization and quantization error enables automated pruning and efficient quantization, improving accuracy and hardware efficiency. • Our two-step training based on pre-trained networks jointly optimizes pruning and quantization, avoiding error accumulation and achieving better compression efficiency and model performance.

Two-phase collaborative model compression training for joint pruning and quantization

Abstract

Journal

Organization