Epistemic Intelligence & Computation Lab

The Epistemic Intelligence & Computation (EPIC) Lab is dedicated to advancing the principles necessary to developing epistemically intelligent systems, i.e. machine learning systems that can recognise, reason about, and communicate the limits of their knowledge. To achieve this, our research spans foundational theory to translational applications, focusing on knowledge-level uncertainty, commonly referred to as epistemic uncertainty.

• Mathematical Foundations of Epistemic Uncertainty: We investigate the mathematical and statistical structures underlying knowledge-level uncertainty. Our work examines formal representation frameworks (e.g., single probability measures, sets of distributions, and higher-order models), principled quantification strategies (e.g., information-theoretic and discrepancy-based characterisations), methods for comparison and geometry (e.g., probability metrics and statistical distances), and procedures for empirical verification (e.g., statistical testing and model criticism). This research draws on tools from probability theory, imprecise probability, Bayesian statistics, information theory, statistical inference and testing, and cooperative game theory.
• Algorithms for Epistemic Uncertainty-Aware Learning: We design and analyse machine learning algorithms that explicitly model, propagate, and exploit epistemic uncertainty across the learning pipeline. Our work focuses on uncertainty-aware predictive modelling (e.g., Gaussian processes, Bayesian neural networks, credal predictors, and conformal prediction methods), structured inference under incomplete or ambiguous information (e.g., causal, counterfactual, and strategic learning), and decision-making under uncertainty (e.g., active learning and Bayesian optimisation). Emphasis is placed on statistical validity, robustness to distributional shift, computational efficiency, and principled uncertainty communication, enabling learning systems to support reliable reasoning and decision processes in open-world settings.
• Epistemic Uncertainty-Aware AI Systems, Frameworks, and Applications: We operationalise our theoretical and algorithmic advances by developing uncertainty-aware AI systems and methodological frameworks for real-world deployment. Our work investigates how epistemic uncertainty can be leveraged to improve reliability, safety, and robustness in complex learning environments, with applications spanning interpretable and explainable models, out-of-distribution detection, robustness under distributional shift, incentive-aware and adversarial learning, AI safety, and foundation / language models. A central objective is the design of systems that can recognise the limits of their knowledge, adapt to novel conditions, and communicate uncertainty in a manner conducive to human oversight and decision-making.