Artificial intelligence systems are increasingly deployed in real-world environments where their predictions influence critical decisions. However, modern AI models often operate under uncertainty, distribution shifts, and unexpected data conditions.
The Trustworthy AI Lab studies how AI systems can operate reliably and safely in real-world environments. Our research focuses on understanding when AI predictions should be trusted, deferred, or rejected, and how AI systems can support reliable decision-making under uncertainty.
Our work integrates machine learning, statistics, and data science to develop AI systems that are not only accurate, but also trustworthy and reliable.
Our research focuses on four core themes:
AI Reliability
Uncertainty
Robust Learning
Human–AI Decision Systems
A central goal of our research is to understand when machine learning models produce reliable predictions and when they fail.
We study how prediction reliability changes under noisy inputs, distribution shifts, and unexpected data patterns. Our work develops methods that allow AI systems to detect unreliable predictions and avoid unsafe decisions.
Reliable AI systems must be able to recognize their own uncertainty.
We investigate methods that allow machine learning models to quantify predictive uncertainty and identify situations where predictions should not be trusted. This includes studying both model uncertainty and data uncertainty and their impact on decision-making.
Machine learning models often encounter conditions that differ from their training environments.
Our research studies how models behave under perturbations, distribution shifts, and adversarial conditions, and develops learning methods that improve robustness and reliability.
In many real-world applications, AI should not make decisions autonomously.
We study decision frameworks where AI systems can defer predictions to human experts when uncertainty is high or predictions are unreliable. Our research explores how AI and humans can collaborate to achieve safer and more reliable decision-making.
Our research focuses on domains where incorrect AI decisions may have significant consequences.
Healthcare AI
Reliable AI systems for medical decision support.
Financial AI
AI systems for financial risk analysis and decision-making.
Physical AI and Autonomous Systems
Trustworthy AI for robotics, autonomous systems, and real-world sensing environments.