Rectifying Classifiers
EurAI joint DL+NMR invited talk
·
Pierre Marquis
CRIL-CNRS/Université d’Artois, Lens, France (homepage)
Dealing with high-risk or safety-critical applications calls for the development of trustworthy AI systems. Beyond prediction, such systems must offer a number of additional facilities, including explanation and verification. The case when the prediction made is deemed wrong by an expert calls for still another operation, called rectification. Rectifying a classifier aims to guarantee that the predictions made by the classifier (once rectified) comply with the expert knowledge. Here, the expert is supposed more reliable than the predictor, but their knowledge is typically incomplete.
Focusing on Boolean classifiers, I will present rectification as a change operation. Following an axiomatic approach, I will give some postulates that must be satisfied by rectification operators. I will show that the family of rectification operators is disjoint from the family of revision operators and from the family of update operators. I will also present a few results about the computation of a rectification operation.