A unified view for unsupervised representation learning with density ratio estimation: Maximization of mutual information, nonlinear ICA and nonlinear subspace estimation