CS Talk
Ieva Daukantas, Stanford University
Host: Ruzica Piskac
Title: Formal Verification of Algorithms Used in Machine Learning: Robust Mean Estimation and Robust Weighted Mean Estimation
Abstract:
Trimming datasets is used as a cleaning or processing technique in many AI systems. It serves the purpose of improving the robustness of AI algorithms, so that they become less vulnerable to data related attacks. Theory states that the outliers in a data set occur with low probability, and so it follows that they can be removed without causing loss of precision in the classification result.
The talk is based on the published paper “Trimming Data Sets: a Verified Algorithm for Robust Mean Estimation” (Ieva Daukantas, Alessandro Bruni, Carsten Schürmann), where we introduce a mechanized proof of robustness of the trimmed mean algorithm. This algorithm has high applicability as a statistical technique and can be used in many complex applications of deep learning. We combine theoretical and practical approaches: the Coq proof assistant is used to formalize the robustness of the trimmed mean algorithm and Python Naïve Bayes experiments to illustrate the applicability in AI systems.
The full publication: “Trimming Data Sets: a Verified Algorithm for Robust Mean Estimation” in PPDP 2021, Authors: Ieva Daukantas , Alessandro Bruni , Carsten Schürmann, https://dl.acm.org/doi/fullHtml/10.1145/3479394.3479412 .
Bio:
Ieva Daukantas is a visiting researcher at Stanford University, USA and a last year PhD at IT University of Copenhagen, Denmark. The research interest areas are formal methods, information security, AI and natural language processing. Before the PhD, she gained 7 years industry experience in leading roles, MSc in Software Development and BSc in Psychology. You can contact her at daukantas@itu.dk .