Say I have 500 variables and I believe those variables can be shown in a 4-dimensional latent representation which I want to learn.
What I have for training is 100K samples, and those samples are coming mainly from 3 unbalanced groups: 1st group has 1K samples, 2nd group has 49K samples, and 3rd group has 50K samples.
Do you think I can learn a meaningful representation by training a (variational) autoencoder with this data? Is there a reason that requires all samples to come from the same distribution? If not, is there a reason that requires balanced classes?