Towards neural networks that provably know when they don't know