The problem is that neural networks have many parameters (weights), that hard to learn. The Jonathan Frankle and Michael Carbin find the pruning technique that can find smaller part of the neural networks that are responsible for resolving current task. The article can found here
This repository contains implementation of that idea on PyTorch and examples with Lenet-300-100 and Conv-Lenet-5.
We can prune 75% of Lenet-300-100 on the MNIST dataset and 25% of Conv-Lenet-5 on SVHN dataset. But it is not end, we didn't search over all different pruning percents for different models due limited resources.