Diversity-guided MLP pruning for efficient large vision transformers

doi:10.1016/j.patcog.2026.114860

Diversity-guided MLP pruning for efficient large vision transformers

2026-09-06

0

PRE

AI

DOI：10.1016/j.patcog.2026.114860

Original

Original request for help

Save

Abstract

En 中文

• MLPs dominate model parameters in large vision transformers. • We propose a Diversity-Guided MLP Pruning to significantly reduce model parameters. • We prune redundant neurons, while preserving weight diversity for better recovery. • Our method reduces parameters and FLOPs by over 57.0% almost losslessly.

Keywords:

MLP pruning

Large vision transformer

Vision language model

Diversity-guided MLP pruning for efficient large vision transformers

Abstract

Journal

Organization

Cited Papers