Neural network pruning using MV regularizer

Proposes a method of neural network pruning by using a strong regularizer called the MV regularizer. The MV regularizer learns a distinct penalty factor attached to each weight by solving a minimization problem over the validation error. After the learning, penalty factors are used to prune network weights one by one by monitoring generalization performance. Experiments using artificial data and real data showed the proposed method worked very well to remove a number of insignificant weights without the loss of generalization