On the Use of Evidence in Neural Networks