Dynamic Hard Pruning of Neural Networks at the Edge of the Internet