Exploring latent spaces with deep autoencoders