Luuk de Jong
Master student at Universiteit Leiden
This project investigates the integration of a reject option in machine learning models to enhance reliability and explainability. By rejecting uncertain predictions, we can mitigate risks associated with low-confidence decisions, meaning the model will be more reliable. The core contribution of this work is the development and evaluation of AutoGASTeN. This automated pipeline leverages generative adversarial stress test networks (GASTeN) to generate realistic images at the decision boundary of two classes. We can create custom datasets with these low-confidence instances for training machine learning models with a reject option.
Keywords: Hyperparameter optimization, Generative AI, Machine learning with a reject option
Scientific area: Artificial Intelligence
Bio: I am a master’s student at the University of Leiden, where I am finishing my master’s thesis together with Carlos Soares and Jan N. van Rijn. My main research interests are Trustworthy AI and Generative AI. I obtained my Bachelor’s degree in Software development from the University of Applied Sciences Utrecht in July 2020. My master’s thesis delved into improving the performance of GASTeN, a framework for generating images on the decision boundary of two classes, and using GASTeN to generate datasets that we can use to train another neural network to reject those instances instead of classifying them.
Visiting period: July 2nd – July 27th at Fraunhofer Portugal AICOS