pruning

Pruning works like cutting off branches that don't help a fruit tree grow fruit — it finds and removes connections and computational parts inside a neural network that have little effect on the result.

A model that has just finished training is like a tree with lots of overgrown branches. Many of its connections don't actually contribute much to producing the answer. Pruning identifies these connections one by one, cuts them out, and adjusts what remains so it still performs close to the original. Just as a trimmed tree becomes lighter, a pruned model becomes smaller and computes faster.

This matters because of where models need to run. A large model can't easily run as-is on small devices like laptops or smartphones. Pruning is one of several techniques used to shrink a model's size so it can run on such devices.