Concept drift detection in image classification

Details

Supervisors

John Lee

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

Despite their power, modern machine-learning tools are sensitive and often fail without warning. To address this issue, there is a need for robust testing frameworks that can detect and quantify the impact of data drift on computer vision classifier in production. During my internship at GSK I was tasked to develop a robust statistical test that measures changes in input distributions truly altering model performance. The central focus of this thesis is the detection of drift using uncertainty evaluation of a model. We explored two main approaches: Monte Carlo Dropout (MCD) and Gaussian Process (GP). Both approaches demonstrated similar performance, effectively estimating uncertainty and enabling successful detection of concept drift based on these uncertainty scores.