Model Pruning
Removing weights or whole structures from a trained network, usually the smallest-magnitude ones first, to shrink it with minimal loss in accuracy.
Think of It Like This
Like trimming dead and spindly branches off an overgrown fruit tree so the remaining limbs produce bigger fruit while taking up less garden space.
Weights below a magnitude threshold get zeroed out after training, on the reasoning that they contributed the least to the network's output. Unstructured pruning removes individual weights for the best accuracy-per-sparsity ratio, while structured pruning removes whole channels or neurons so ordinary hardware can realize the memory and speed savings without special sparse-matrix support.
Fine-tuning the surviving weights after pruning is standard practice, since the remaining weights were trained assuming their now-removed neighbors were still contributing.