Compressing neural network by tensor network with exponentially fewer variational parameters