Pruning removes weights or entire channels that contribute little to the output, producing sparse structures to reduce computation.