Laplacian Autoencoders for Learning Stochastic Representations

The supplementary material is organized as follows. Second, we discuss the Laplace approximation and the "mean shift" issue encountered Third, we elaborate more on the difference between related works and our proposed method. For validation, we sampled 5000 data points randomly from the training sets. We keep the same dropout rate during testing. We use a diagonal approximation of the hessian in all experiments.