A Geometric Perspective on Variational Autoencoders