Recovering from Random Pruning: On the Plasticity of Deep Convolutional Neural Networks