2026 Seminars

This presentation has two parts. In the first, I will discuss some of the challenges of working with adaptive designs. These are experimental designs in which the next input may depend on the data collected up to that point, breaking independence between observations. As a result, the large-sample theory of adaptive designs is significantly more difficult than in the i.i.d. case, and asymptotic normality must often be verified on a case-by-case basis. I will present open problems concerning the asymptotic properties of designs currently used in the energetics industry and research, developed to test the sensitivity of explosives.

The second part of the talk focuses on applications of TabPFN. Introduced in January last year, with a major update in October, TabPFN is a foundation model for tabular data that outperforms state-of-the-art methods such as XGBoost on many regression and classification tasks. Unlike traditional machine learning approaches, it requires no fine-tuning or additional training: TabPFN is pre-trained on an enormous corpus of synthetic datasets designed to capture nonlinear relationships in tabular problems. In addition to outlining ideas for future use, I will present one application in which we use TabPFN to improve estimates of conditional Shapley values in explainable AI.

About the speaker: Dennis Christensen is a researcher at the Norwegian Defence Research Establishment (FFI), visiting the University of Auckland from early January until mid-April. His research focuses on statistical aspects of sensitivity testing of energetic materials., with particular focus on explosive remnants of war and dumped ammunition. He completed his PhD at the University of Oslo in 2024.

Abstract: We propose a novel modeling framework for time-evolving networks allowing for long-term dependence in network features that update in continuous time. Dynamic network growth is functionally parameterized via the conditional intensity of a marked point process. This characterization enables flexible modeling of both the time of updates and the network updates themselves, dependent on the entire left-continuous sample path. We propose a path-dependent nonlinear marked Hawkes process as an expressive platform for modeling such data; its dynamic mark space embeds the time-evolving network. We establish stability conditions, demonstrate simulation and subsequent feasible likelihood-based inference through numerical study, and illustrate the methodology with an application to conference attendee social network data. The resulting methodology serves as a general framework that can be readily adapted to a wide range of network topologies and point process model specifications.