Laplacian Autoencoders for Learning Stochastic Representations