As an expert in your field, you know exactly which problems are worth solving.

You know what to solve, but not necessarily the best way to do so...

Stuck on a problem in your research?

Do you feel like you are doing repetitive steps over and over again?

Do you feel like there is probably a smarter way to do things?

As an expert in your field, you tackle multivariate distributions daily.

You just know how to efficiently reach a solution.

Do you have a great machine learning method only demonstrated on toy data?

Do you need a real-world problem to test your novel solutions?

Drs. Arthur Thuy - Uncertainty Quantification in Neural Networks: From Trustworthy Decision-Making to Efficient Data Collection

Abstract: Neural networks (NNs) are increasingly prominent in engineering applications and are deployed across a wide range of data modalities, including tabular, textual, and image data. While NNs demonstrate strong predictive performance, they are also known to be confidently wrong under distribution shifts, where the test data diverge from the training distribution. In such scenarios, model performance can degrade without clear warning. Uncertainty quantification (UQ) provides a principled approach to address this issue and indicates when outputs should—or should not—be trusted. This talk explores three key roles of UQ in enhancing model robustness and usability. First, we position UQ as a tool for model explainability, connecting it to downstream decision-making through classification with rejection. This approach allows uncertain predictions to be deferred to human experts, thereby reducing the number of misclassifications. Second, we examine the role of UQ in active learning, where it guides the efficient selection of informative samples for annotation, minimizing labeling effort while preserving model accuracy. Finally, we discuss methods for obtaining high-quality UQ from efficient neural network ensembles, reducing computational costs and increasing the practicality of UQ in real-world applications.

Biography: Arthur Thuy is a doctoral researcher in the research group of Prof. dr. Dries Benoit at Ghent University and a PhD fellow of the Research Foundation –Flanders (FWO). He is also affiliated with the CVAMO Lab at Flanders Make. His research focuses on supporting business decision-making with uncertainty quantification in neural networks, communicating when a model's output should (not) be trusted. His work spans applications in both the manufacturing industry and educational data mining. His research has been published in the European Journal of Operational Research, Annals of Operations Research, and presented at ECML-PKDD.

Prof. dr. Carl Henrik Ek - Approximate Bayesian Inference of Composite Functions

Abstract:

Over the last decade, machine learning methods built on composite functions have lead to a remarkable increase in predictive power. Building complex structures from compositions of simple mathematical objects has led to a surprising generalisation of success across many domains. The aspiration of Bayesian deep learning is to combine the predictive advantages of deep neural networks with the theoretical justifications of the Bayesian framework, hoping to achieve the benefits of both. Success in achieving the promises of a Bayesian treatment of deep neural networks has been so elusive that it raises the question of whether there is something inherent in their structure that renders it numb to a Bayesian approach.

In this talk, I will first highlight the inherent structures in the compositional models that makes Bayesian approximate inference challenging. We will provide a characterization of why Bayesian learning has not succeeded, showing that for deep learning to harvest the benefits of a Bayesian treatment, this structure needs to be reflected in the approximate posterior. We will then discuss how we can use tools from differential geometry to characterise the structures of these parameter spaces to design efficient computational models to circumvent these issues and design tailored approximate inference schemes.

Biography:

Carl Henrik Ek is a Professor of Statistical Learning at the University of Cambridge and a Professor at Karolinska Institute in
Stockholm. He is the co-director for the UKRI AI Centre for Doctoral Training in Decision Making for Complex Systems which is a collaboration between University of Cambridge and the University of Manchester. He is also a fellow and Director of Studies in Computer Science at Pembroke College, Cambridge. 
Carl Henriks research interests is focused on machine learning and in specific probabilistic models. He has worked extensively in Bayesian non-parametrics and have developed techniques for latent variable models using Gaussian processes and how to perform efficient inference in such models. He has also worked extensively on applications of these techniques in a range of different domains, from renewable energy to astrophysics. Over the last few years his main focus area have been applications within health and medicine.