Tiltrædelsesforelæsninger ved Markussen og Clemmensen
Inaugural lectures by Bo Markussen and Line Katrine Harder Clemmensen.
Program:
14.00-14.45 Line Katrine Harder Clemmensen: A statistical modelling journey for low resource applications in high dimensions
14.50-15.35 Bo Markussen: Evidence from Small Datasets and the Role of the Null Hypothesis
15.40-16.30 Reception in the South End of Vandrehallen
Line Katrine Harder Clemmensen: A statistical modelling journey for low resource applications in high dimensions
Abstract: This talk presents modern statistical modelling strategies and their development in the context of low resources and high dimensions. I will discuss data or domain-driven priors and constraints, data representativity, evaluation strategies, and self-explainability concepts useful when developing safe artificial intelligence tools. From regularized statistical models to deep learning strategies with constraints, the talk will consider the progress in modelling strategies with applications in spectral data analysis e.g., food safety/control or astrophysics, and multi-modal data analysis with mental health applications.
My aim is to take you on a journey through my research career and, through this, reflect on the development of modeling strategies over the years.
Bo Markussen: Evidence from Small Datasets and the Role of the Null Hypothesis
ABSTRACT: In contrast to the burgeoning world of big data, predictive modelling, and AI, I will discuss two key virtues of classical statistics: small datasets and hypothesis testing. A criminal court case serves as an archetypal metaphor for hypothesis testing. A fundamental principle in court is that the accused is presumed innocent until proven guilty. In the common statistical setup, innocence corresponds to the null hypothesis, and compelling evidence against it is indicated by observing a small p-value. In this framework, guilt is established by rejecting the null hypothesis of innocence. I will provide a real-life example from my own work in the Data Science Laboratory to illustrate this concept. Situations where the accused is guilty, but the forensic evidence is likely to be insufficient to reject the null hypothesis, are referred to as low statistical power. In such cases, one may look for evidence against weaker versions of the null hypothesis. Similarly, alternative formulations of the null hypothesis may be required to establish evidence for other properties. I will share examples of both situations from recent collaborations in the Data Science Laboratory.
My aim is to present these ideas in an accessible manner, with the hope that non-statisticians and even non-mathematicians will be able to follow and appreciate the concepts. If time permits, I will conclude my talk with a brief reflection on the position and purpose of the Data Science Laboratory at UCPH-SCIENCE.