Inference
Magnitude pruning
2015ActiveUpdated: 22 September 2026Published
Key
innovation
Making a weight's absolute value the criterion for removal — a simple, strong baseline.
Category
Inference
Abstraction level
Building block
Operation level
Post-training
Use cases
Model compressionSparsityLottery ticket hypothesis
How it works
A magnitude threshold is set and weights below it are zeroed; optionally the model is fine-tuned iteratively.
Problem solved
Neural networks are over-parameterized; redundant weights must be cheaply removed without much quality loss.