Cooperation in the Latent Space: The Benefits of Adding Mixture Components in Variational Autoencoders