Movement Pruning: Adaptive Sparsity by Fine-Tuning

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.

arXiv comment14 pages, 6 figures, 3 tables. Published at NeurIPS2020. Code: \url{huggingface.co/mvp}

References in corpus (17)

• Distilling the Knowledge in a Neural Network
• HuggingFace's Transformers: State-of-the-art Natural Language Processing
• Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
• Compressing Deep Convolutional Networks using Vector Quantization
• To prune, or not to prune: exploring the efficacy of pruning for model compression
• SNIP: Single-shot Network Pruning based on Connection Sensitivity
• Q8BERT: Quantized 8Bit BERT
• The State of Sparsity in Deep Neural Networks
• Distilling Task-Specific Knowledge from BERT into Simple Neural Networks
• Reducing Transformer Depth on Demand with Structured Dropout
• Sparse Networks from Scratch: Faster Training without Losing Performance
• TinyBERT: Distilling BERT for Natural Language Understanding
• Global Sparse Momentum SGD for Pruning Very Deep Neural Networks
• Training with Quantization Noise for Extreme Model Compression
• Playing the lottery with rewards and multiple languages: lottery tickets in RL and NLP
• Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Bias
• Reweighted Proximal Pruning for Large-Scale Language Representation

Cited by in corpus (17)

• Layer-adaptive sparsity for the Magnitude-based Pruning
• Only Train Once: A One-Shot Neural Network Training And Pruning Framework
• Text-Conditioned Transformer for Automatic Pronunciation Error Detection
• Parameter-Efficient Transfer Learning with Diff Pruning
• A Survey on Self-supervised Pre-training for Sequential Transfer Learning in Neural Networks
• I-BERT: Integer-only BERT Quantization
• PARP: Prune, Adjust and Re-Prune for Self-Supervised Speech Recognition
• Pixelated Butterfly: Simple and Efficient Sparse training for Neural Network Models
• TX-Ray: Quantifying and Explaining Model-Knowledge Transfer in (Un-)Supervised NLP
• Plug-Tagger: A Pluggable Sequence Labeling Framework Using Language Models
• Deep Neural Compression Via Concurrent Pruning and Self-Distillation
• SparseDNN: Fast Sparse Deep Learning Inference on CPUs
• ESPN: Extremely Sparse Pruned Networks
• Doping: A technique for efficient compression of LSTM models using sparse structured additive matrices
• When deep learning models on GPU can be accelerated by taking advantage of unstructured sparsity
• Pruning Attention Heads of Transformer Models Using A* Search: A Novel Approach to Compress Big NLP Architectures
• How Well Do Sparse Imagenet Models Transfer?