Dirichlet Pruning for Neural Network Compression