Yavuz Durmazkeser

Despite its strengths in fast, automated decision-making, data-driven learning remains vulnerable to real-world perturbations and adversarial attacks. This PhD thesis aims to certifiably enhance the adversarial and distributional robustness of machine learning models by integrating knowledge and reasoning. Its focus will be on two pillars: Explicit and Implicit Reasoning. For explicit reasoning, we will combine rule-based methods with data-driven techniques to enhance model robustness. For implicit reasoning, we will leverage knowledge bases to enrich and factually ground existing data sources. Our single overarching goal is to mimic human-in-the-loop learning for more transparent, secure and trustworthy AI systems.