Sparsity in Deep Learning: Pruning + growth for efficient inference and training in neural networks