Uniform Transformation: Refining Latent Representation in Variational Autoencoders