Keynote Speakers

Prof. Amara Okonjo School of Informatics, University of Edinburgh · Day 1, 09:30 — Beveridge Hall

Auditing what we cannot see: evaluation under partial access

Most external audits of deployed models are conducted without weights, training data or logs. This talk sets out what can and cannot be established from query access alone, formalises the resulting identification limits, and proposes a minimal disclosure regime under which audit conclusions become falsifiable rather than merely plausible.

Okonjo leads the Accountable Systems Group in Edinburgh and has advised three national algorithmic-transparency reviews.

Prof. Henrik Lindqvist Technology, Policy and Management, TU Delft · Day 1, 16:30 — Beveridge Hall

Causal foundations for accountable decision systems

Fairness and robustness criteria stated in purely predictive terms are unstable across deployment shifts. Lindqvist argues for stating obligations causally — over interventions and recourse rather than over correlations — and reviews five years of field studies in public-sector allocation where causal specification changed the decision, not only the diagnostic.

Chair in Decision Systems and Public Values at TU Delft; formerly head of analytics at the Netherlands Court of Audit.

Dr. Priya Raghunathan Centre for Data Provenance, Imperial College London · Day 2, 09:15 — Beveridge Hall

Provenance at scale: documenting corpora no one has read

Documentation practices designed for curated datasets fail on web-scale corpora. This talk presents measurement work across sixteen public training sets, the systematic ways licence and consent metadata degrade during aggregation, and a sampling protocol that yields defensible corpus-level claims at tractable cost.

Her group builds provenance infrastructure used by three national research clouds.

Prof. Wei-Lin Chen Institute of Data Science, National University of Singapore · Day 2, 15:30 — Beveridge Hall

Monitoring after deployment: drift, feedback and analytics that survive contact

Offline evaluation predicts deployed performance poorly wherever a model's own outputs shape the data it next observes. Chen surveys feedback-induced drift in credit, health triage and content ranking, and sets out monitoring designs that keep post-deployment claims interpretable.

Directs the Deployment Analytics Lab at NUS and chairs the ACM SIGMOD reproducibility committee.