Programme

What if we train a model to classify dogs and cats, but it is later tested with an image of a human? Generally the model will output either dog or cat, and has no ability to signal that the image contains no class that it can recognize. Similarly for systems like ChatGPT, they are not able to convey their own uncertainty, which means they cannot tell the user the confidence on their answer. This is because classical neural networks do not contain ways to estimate their own uncertainty (so called epistemic uncertainty), and this has practical consequences for the use of these models, like safety when cooperating with humans, autonomous systems like robots, computer vision systems, healthcare applications, and other uses that require reliable uncertainty quantification estimates. In this talk we will cover the basic concepts of uncertainty quantification, how to train machine learning models with uncertainty, Bayesian neural networks, how to evaluate them, and related benchmarks and evaluation metrics. This field combines concepts from Statistics into Machine Learning models, and the talk will cover methods from the state of the art, but only requiring basic background in Machine Learning and Statistics.

In this research, we employ a particle-based Bayesian inversion algorithm to address the practical challenge of obtaining less conservative bounds for the operation of underground, medium-voltage power cables. The greater objective is to work towards a real-time enhancement of the efficiency and performance of underground cables in the Netherlands. In this project, we aim to deduce the thermal properties of the surrounding soil from an observed thermal profile, as well as quantifying the reliability (or uncertainty in) the obtained estimates. We obtained the thermal profile of the cables by in-the-field measurements in the Netherlands, collected in the last few months. This study is being conducted in collaboration with Alliander, a prominent grid operator responsible for a segment of the Dutch electricity network.

I will discuss a methodology to calibrate mathematical models for tumour cell dynamics that take into account the effects of their surrounding microenvironment. The adoption of a Bayesian framework allows to quantify the uncertainty in the parameter estimates as well as possible correlations in the parameters. We will also see how the Bayesian framework can be useful to compare different models. I will present calibration results using different in-vitro laboratory data, where the environmental factors are either nutrient deprivation or oxygen level, tissue stiffness and the presence of chemotherapeutic drugs. The analysis of the calibration results will allow to shed some light on the underlying biological mechanisms affecting tumour progression. This is joint work with Sabrina Schoenfeld (TU Munich), Christina Kuttler (TU Munich), Alican Ozkan (Harvard University) and Marissa Nichole Rylander (UT Austin).