Definition
Plain language
Deleting some of a neural network's weights to shrink it or change its behavior.
As stated in the literature
Removing parameters from a trained network, often the smallest-magnitude weights; in FloatDoor, pruning roughly 10% of weights destroys the backdoor while leaving benchmark scores nearly unchanged.
Also called: prune, magnitude pruning
Why it matters: It can shrink a model or strip out unwanted behavior with little cost to accuracy, making it both an efficiency and a security tool.
For example, deleting the smallest 10% of a network's weights can wipe out a hidden backdoor while leaving its normal performance almost untouched.
Heard on the show
“… cluster would have more and more keys in it, the balls would get bigger to enclose them, and your pruning power would just evaporate. …”Episode 036 — Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.