The paper presents an approach of increasing classification accuracy by transforming the low confident predictions to the space of the inputs that are easier to classify. Well-written paper contrasting with related work, though missing some of the other relevant work. The limitations of the work is the simplicity of the datasets used. Would have liked to see more exposition on the out-of-distributions that are being modeled (and why de-noising methods won't work), especially since the experiments are performed on MNIST.