My research is part of an interdisciplinary, DFG-funded research unit on (deep) anomaly detection for chemical process data1.
In my work, I leverage techniques from the domain of Formal Methods to improve machine learning (models):
I develop efficient methods to verify, i.e., to prove formally the safety and reliability of (deep) neural networks used in anomaly detection.
Furthermore, I investigate how logical specifications can be integrated into explainable AI (XAI) methods to overcome their lack of rigor and provide more concise explanations.
Moreover, I develop new techniques for training neural networks to make them safe by construction, i.e., one can ensure that a neural network satisfies a set of correctness properties at the end of training.
Adopting Formal Methods will allow us to prove the correctness, safety, and reliability of AI systems and provide better explanations.
This will ultimately aid the process of building trustworthy AI.
