Deep neural networks are large models and pruning has become an important part of ML product pipelines, making models small while keeping their performance high. However, the classic pruning method, Magnitude Pruning, is suboptimal in models that are obtained by transfer learning. This paper proposes a solution, called Movement Pruning and shows its superior performance.
OUTLINE:
0:00 – Intro & High-Level Overview
0:55 – Magnitude Pruning
4:25 – Transfer Learning
7:25 – The Problem with Magnitude Pruning in Transfer Learning
9:20 – Movement Pruning
22:20 – Experiments
24:20 – Improvements via Distillation
26:40 – Analysis of the Learned Weights
Paper:
Code:
Abstract:
Magnitude pruning is a widely used strategy for reducing model size in pure supervised learning; however, it is less effective in the transfer learning regime that has become standard for state-of-the-art natural language processing applications. We propose the use of movement pruning, a simple, deterministic first-order weight pruning method that is more adaptive to pretrained model fine-tuning. We give mathematical foundations to the method and compare it to existing zeroth- and first-order pruning methods. Experiments show that when pruning large pretrained language models, movement pruning shows significant improvements in high-sparsity regimes. When combined with distillation, the approach achieves minimal accuracy loss with down to only 3% of the model parameters.
Authors: Victor Sanh, Thomas Wolf, Alexander M. Rush
Links:
YouTube:
Twitter:
BitChute:
Minds:
source