Peter Flach is a Professor of Artificial Intelligence at the University of Bristol, School of Computer Science. His expertise in mining highly structured data and in the evaluation and improvement of machine learning models has positioned him as an internationally leading researcher in these areas. Peter is the Director of the UK Research and Innovation (UKRI) Centres for Doctoral Training (CDTs) in Interactive Artificial Intelligence and Practice-Oriented Artificial Intelligence. 

Peter focuses on data-driven and knowledge-driven approaches of: machine learning, data science and AI (human-centered-AI interaction and responsible AI).

Keynote Details

Title: The Why and How of Classifier Calibration
Abstract: A well-calibrated classifier correctly quantifies the level of uncertainty or confidence associated with its instance-wise predictions. This is essential for critical applications, optimal decision making, cost-sensitive classification, and for some types of context change. I will survey the main concepts and methods, including proper scoring rules and other evaluation metrics, visualisation approaches, and post-hoc calibration methods for binary and multiclass classification.

Thomas is the founder of PyMC Labs, a consulting company that provides transformative innovation for businesses with a combination of cutting-edge probabilistic AI and world-class expertise.  The PyMC Labs are the creators of PyMC, the leading open-source software for statistic and probabilistic AI modeling. 

Keynote Details

Title: Explainable Modeling, Agentic Workflows, and Decision-Grade Inference for Pharma
Abstract: Bayesian modeling is the most effective path when answers must be explainable, auditable, and decision-grade. I will show why, using a recent pharma case: a Bayesian hierarchical Gaussian-process framework for rodent home-cage activity in neurodegeneration. The model captured circadian structure and long-term decline, produced time-resolved effect sizes, recovered known phenotype traits, and improved power and sample-size planning—without collapsing data into coarse day/night bins. This is what generative, causal, hierarchical Bayes delivers: integration over uncertainty, explicit mechanism, and actionable “what-if” queries.

The barrier has been usability. Building robust Bayesian models is hard. That constraint is changing. Agentic systems now make these methods practical at scale: agents specify candidate models in a PPL (e.g., PyMC/Stan), run inference, perform posterior predictive checks and SBC, probe sensitivity to priors and likelihoods, test causal queries, cross-validate, and auto-generate human-readable reports with effect sizes, practical-equivalence probabilities, and study-design simulations. Every step is logged with an audit trail and reproducible artifacts.

The result is a governed workflow that pairs black-box ML for prediction with Bayes for interpretation and intervention, spanning RCTs, real-world data, and digital biomarkers. This keynote will outline the principles, show the Roche home-cage case end-to-end, and map how agentic Bayesian pipelines turn complex pharma data into transparent, defensible decisions.