Gravar-mail: Efficient probabilistic inference in generic neural networks trained with non-probabilistic feedback