Keynote Speakers

Julie Josse

Inria-Inserm Premedical team, Montpellier, France

Title

Beyond the Average Doctor: Uncertainty-Aware AI for Clinical Triage

Abstract

This talk traces the development of Traumatrix, a decision support tool embedded in ambulances to optimize triage, patient care, and resource allocation for severe trauma. What initially appeared to be a straightforward supervised learning problem quickly revealed the many challenges of translating AI from the lab to an upcoming nationwide cluster randomized controlled trial. While predictive models can achieve excellent performance, deploying them safely in the chaotic, time-critical setting of trauma care requires more than accuracy. This talk shows how uncertainty emerges at every stage of the AI lifecycle and presents practical strategies to manage it. We will examine how uncertainty arises from incomplete data and imperfect clinical labels, how to quantify predictive uncertainty, when AI should defer to clinicians, and how to account for human–AI interaction. Finally, the talk explores the challenge of evaluating models against historical clinical decisions that may not reflect true patient benefit, showing how a causal framework can guide both model development and the evaluation of AI systems in real-world clinical practice.

Bio

Julie Josse is a senior researcher at Inria, leading the PreMeDICaL team with Inserm. Their work advances personalized medicine through causal learning and federated methods that preserve data confidentiality, aiming to accelerate targeted therapies and decision-support tools with quantified uncertainty. Julie Josse’s expertise includes missing data, causal inference, and machine learning on multi-source, multi-modal health data, enhancing decisions in respiratory disease, oncology, and fertility. She led the Traumatrix project, creating decision-support tools for ambulance trauma care optimization. Prior to Inria, she was a professor at Ecole Polytechnique (Institut Polytechnique de Paris – IP Paris), directing the Data Science for Business Master’s with HEC Paris. She has also held visiting roles at Stanford University and Google Brain Paris and has been recognized with honors such as the Inria–French Academy of Sciences Young Researchers Prize.

Lester Mackey

Microsoft Research New England, United States

Title

Compressing Uncertainty in Artificial Intelligence

Abstract

This talk will introduce three new tools for summarizing a probability distribution more effectively than independent sampling:

Given an initial n point summary (for example, from independent sampling or a Markov chain), kernel thinning finds a subset of only square-root n points with comparable integration error.
If the initial summary suffers from biases due to off-target sampling, tempering, or burn-in, Stein kernel thinning simultaneously compresses and corrects the summary for bias.
Compress++ accelerates distribution compression by converting quadratic-time thinning algorithms into near-linear-time algorithms with comparable error.

These tools are especially well-suited for tasks that incur substantial downstream costs per summary point like organ simulation in computational cardiology and long-context attention in transformers.

Bio

Lester Mackey is a Senior Principal Researcher at Microsoft Research, where he develops machine learning methods, models, and theory for large-scale learning tasks driven by applications from meteorology, healthcare, and the social good. Lester moved to Microsoft from Stanford University, where he was an assistant professor of Statistics and, by courtesy, of Computer Science. He earned his PhD in Computer Science and MA in Statistics from UC Berkeley and his BSE in Computer Science from Princeton University. He co-organized the second place team in the Netflix Prize competition for collaborative filtering; won the Prize4Life ALS disease progression prediction challenge; won prizes for temperature and precipitation forecasting in the yearlong real-time Subseasonal Climate Forecast Rodeo; and received best paper, outstanding paper, and best student paper awards from the ACM Conference on Programming Language Design and Implementation, the Conference on Neural Information Processing Systems, and the International Conference on Machine Learning. He is a 2023 MacArthur Fellow, a Fellow of the Institute of Mathematical Statistics, a Fellow of the American Statistical Association, an elected member of the COPSS Leadership Academy, and the recipient of the 2023 Ethel Newbold Prize and the 2025 COPSS Presidents' Award.

Julia Stoyanovich

New York University, United States

Title

Uncertainty in the Data Engineering Pipeline: From Imputation to Explanation

Abstract

Machine learning pipelines are commonly evaluated on model accuracy, yet much of the uncertainty that shapes real-world outcomes originates not in the model itself but in the data engineering decisions that precede it. This talk examines two manifestations of that upstream uncertainty and their downstream consequences for trustworthy AI. The first concerns missing value imputation. Real-world missingness is rarely as clean as Rubin's classic MCAR/MAR/MNAR taxonomy assumes: datasets exhibit multi-mechanism missingness and missingness shift between training and deployment. Drawing on the Shades-of-Null evaluation suite and the VirnyFlow framework, I will show how imputation strategies induce variation in predictive accuracy, model stability, and fairness across demographic groups, and that these dimensions do not move together. No single best imputer exists; the right choice depends on the missingness regime, the model class, and the stakeholder context. The second manifestation concerns post-hoc explanations. SHAP-based feature attributions are increasingly used to justify decisions and satisfy regulatory requirements, yet they are surprisingly fragile. Routine transformations such as bucketizing a continuous feature or encoding a categorical one can drastically shift which features are deemed most important, opening the door to inadvertent or adversarial manipulation. Compounding this, explanation multiplicity is the phenomenon of substantial disagreement in feature attributions across repeated runs of the same pipeline, with the model and input held fixed. It is widespread even for high-confidence predictions, and is systematically masked by commonly used metrics. Together, these results argue for treating data engineering as a first-class source of uncertainty in responsible ML, and for evaluation frameworks that make this uncertainty visible, measurable, and governable.

Bio

Dr. Julia Stoyanovich is the Institute Associate Professor of Computer Science and Engineering, Associate Professor of Data Science, and Director of the Center for Responsible AI (https://r-ai.co) at New York University. Her goal is to make “responsible AI” synonymous with “AI.” She pursues this goal through academic research, education, technology policy, and public engagement, regularly speaking about both the benefits and the risks of AI. Her research spans data management and AI systems, as well as the ethics and governance of AI. Julia holds an M.S. and Ph.D. in Computer Science from Columbia University, and a B.S. in Computer Science and in Mathematics & Statistics from the University of Massachusetts Amherst. She is a recipient of the Presidential Early Career Award for Scientists and Engineers (PECASE) and is a Senior Member of the Association for Computing Machinery (ACM).

Last updated: July 27, 2026 13:07 (UTC)

Sponsors: