Bunse/2021a: Certification of Model Robustness in Active Class Selection

Bibtype Inproceedings
Bibkey Bunse/2021a
Author Bunse, Mirko and Morik, Katharina
Ls8autor Bunse, Mirko
Morik, Katharina
Title Certification of Model Robustness in Active Class Selection
Booktitle Europ. Conf. on Mach. Learn. and Knowledge Discovery in Databases (ECML-PKDD)
Publisher Springer
Abstract Active class selection provides machine learning practitioners with the freedom to actively choose the class proportions of their training data. While this freedom can improve the model performance and decrease the data acquisition cost, it also puts the practical value of the trained model into question: is this model really appropriate for the class proportions that are handled during deployment? What if the deployment class proportions are uncertain or change over time? We address these questions by certifying supervised models that are trained through active class selection. Specifically, our certificate declares a set of class proportions for which the certified model induces a training-to-deployment gap that is small with a high probability. This declaration is theoretically justified by PAC bounds. We apply our proposed certification method in astro-particle physics, where a simulation generates telescope recordings from actively chosen particle classes.
Note To appear
Year 2021
Url https://2021.ecmlpkdd.org/wp-content/uploads/2021/07/sub_598.pdf

  • Privacy Policy
  • Imprint