Weakly Supervised Learners for Correction of AI Errors with Provable Performance Guarantees

We  present  a  new  methodology  for  handling  AIerrors by introducing weakly supervised AI error correctors witha prioriperformance guarantees. These AI correctors are auxiliarymaps  whose  role  is  to  moderate  the  decisions  of  some  previouslyconstructed underlying classifier by either approving or rejectingits  decisions.  The  rejection  of  a  decision  can  be  used  as  a  signalto  suggest  abstaining  from  making  a  decision.  A  key  technicalfocus  of  the  work  is  in  providing  performance  guarantees  forthese new AI correctors through bounds on the probabilities ofincorrect decisions. These bounds are distribution agnostic anddo not rely on assumptions on the data dimension. Our empiricalexample illustrates how the framework can be applied to improvethe performance of an image classifier in a challenging real-worldtask  where  training  data  are  scarce.