On the detrimental effect of invariances in the likelihood for variational inference