A pruning method identifies weights, units or larger structures that can be removed under a chosen criterion. The resulting model may require further adaptation and evaluation to recover or preserve useful behavior.
Smaller storage does not automatically translate into faster execution on every device. Hardware support, the structure of the removed components and changes in task quality determine the practical benefit.
ELI5
Pruning means taking selected parts out of a model to make it lighter. The hard part is deciding what can be removed without losing capabilities that matter.
For example, a team might remove parts that appear to contribute little on its test tasks, then compare the smaller model on both familiar and new cases. Good coding results would not prove that its language abilities were unchanged.
Is pruning the same as quantization?
No. Pruning removes components, while quantization represents numerical values with reduced precision.
Does pruning always improve speed?
No. The runtime and hardware must be able to benefit from the resulting structure.
