I'm working with VAE and I got that problem of balacing KL-Divergence and Reconstruction loss, and I'm wondering why use a Gaussian distribution to model the latent space. Is it because we assume that out input follows a Gasussian distribution? My point is that I only get a decent Reconstruciton Loss if I set betha around 0.0001, so I understand that my latent space doesn't look like a Gaussian anymore. And if I set betha to 1, I undestand that my latent space is similar to a Gaussian but the Reconstruction Loss is too high, so I assume that my data cannot be packed into a Gaussian. So... why do we use always a Gaussian? How should I know which distribution is better? Does it depends on the distribution of the input data?
Thank you so much!