The article proposes that the neural network itself is already 'clean', and what is not clean is the way it is observed, with no need to be 'fixed' by external intervention.