On the detrimental effect of invariances in the likelihood for variational inference Richard Kurle Yuyang Wang Jan Gasthaus AWS AI Labs