Trustworthy AI: Exploring Reliability, Transparency and Adaptability of Machine Learning Systems.

Machine learning is becoming deeply embedded in everyday decision making, from digital assistants to safety-critical infrastructure. As these systems grow more pervasive, trust becomes essential. Trustworthy machine learning rests on three pillars: reliability, adaptability, and transparency. Reliability ensures that models behave predictably under distribution shift, resist spurious correlations, and remain resilient to rare events. Adaptability enables systems to adjust to evolving environments, scarce or shifting data, and the feedback loops that naturally arise in real-world deployment. Transparency provides interpretable reasoning and allows for meaningful human oversight. Together, these qualities move trust beyond accuracy alone and toward machine learning systems that are stable, auditable, and aligned with human goals and constraints, even under uncertainty and change. This thesis treats human interaction, control considerations, and structural inductive biases as primary design principles, and is organized into three main parts. First, we incorporate human feedback into anomaly detection models, enabling them to adapt to evolving definitions of abnormality and remain robust against misalignment with expert intent. Second, we investigate the use of safe and interpretable reinforcement learning for safety-critical scenarios, where robustness and reliability of control policies are paramount. Finally, we explore how to encode desirable properties directly into neural network architectures, providing guarantees that improve robustness, transparency, and trustworthiness. Together, these contributions advance the design of machine learning systems that go beyond accuracy, going towards Machine Learning systems that can be safely and confidently deployed in real-world decision making.