Stacked variational denoising autoencoder