Identifying Untrustworthy Predictions in Neural Networks by Geometric Gradient Analysis