Keynote speakers

 

Dr. Paul-Gauthier Noé

Paul-Gauthier is a CNRS researcher in Computer Science in the Laboratoire d'Informatique et des Systèmes in Marseille. Before that, he worked as a postdoctoral researcher at INRIA Grenoble and was a PhD student at Avignon University. He discovered CoDA in Avignon while working on privacy in speech technology and calibration, realising that the Aitchison geometry of the simplex can be used to tackle the questions he faced regarding the calibration of probabilities and likelihoods in multiclass settings.

In 2023, he received the best thesis award from the Association Francophone de la Communication Parlée (AFCP), a special interest group of the International Speech Communication Association (ISCA). At CoDaWork2024, he received the best poster contribution award for his work on the calibration of evidence functions and discriminant analysis.

Today, his research interests still include calibration and privacy in speech technologies, but also include statistical learning in general, explainability in machine learning, and of course the use of CoDA principles within these contexts.

 

 

 

 

Patricia Genius Serra

I am psychologist and statistician, currently working as a fourth-year PhD student at the Barcelonaβeta Brain Research Center in Barcelona.My research focuses on brain imaging genetics, applying bioinformatics methodologies to explore the role of genetic factors in early AD brain changes. My work is conducted within the Genetic Neuroepidemiology and Biostatistics group, led by Dr. Natalia Vilor-Tejedor and co-supervised by Dr. Juan D. Gispert.

Additionally, I serve as an assistant professor at Vic University-Central University of Catalonia, where I teach "Programming" in the Master in Omics Data Analysis, master that I completed in 2021 to deepen my expertise in bioinformatics. In March 2025 I started a research stay at the Genetics Department of Amsterdam UMC, under the supervision of Dr. Sven van der Lee, where I contributed to projects focused on the genetic architecture of neurodegenerative diseases. In January 2026, I will start a new position as a post-doctoral researcher in the Amsterdam UMC, under the supervision of Dr. Betty Tijms, to explore the proteomic profile associated with cerebrovascular damage in individuals diagnosed with AD and vascular dementia.

My background allows me to integrate statistical, bioinformatics, and psychological perspectives to contribute to a better understanding of neurodegenerative disorders.

 

Dealing with multiclass and multiple hypotheses problems with the Aitchison geometry of the simplex

Discrete probability distributions are key elements in statistics and machine learning. In particular, classification models and probabilistic prediction produce probability distributions over a finite set of possible classes or hypotheses. While the Aitchison geometry of the simplex has been actively promoted for the analysis of compositional data, and despite the neat result that perturbation is the Bayes' rule, the use of the Aitchison geometry on the probability simplex has been less discussed in the context of statistical inference and machine learning.
In this talk, we will argue that the linear structure provided by the Aitchison geometry of the simplex allows us to tackle multiclass problems going beyond the standard one-vs-one and one-vs-rest approaches. In particular, we will lean on two practical examples. First, we will see how the Aitchison geometry can be used to extend results on the calibration of probability distributions and likelihood functions from the binary to the multiclass case. Secondly, we will see how the Aitchison geometry can be used to extend the concept of Shapley value to the simplex for explaining probabilistic predictions. With these two practical yet very distinct examples, we aim to demonstrate the wide applicability and relevance of the Aitchison geometry on the probability simplex in statistics and machine learning.

 

 

 

 

CoDa in neuroimaging studies: exploring its application to identify early brain changes in Alzheimer's Disease

In neuroimaging studies, common multivariate methods rely on pairwise covariances between regions, overlooking the broader compositional and bounded nature of the brain. In neurodegenerative diseases, such as Alzheimer’s disease (AD), several neuroimaging measures have been widely described as intermediate phenotypes, which are measurable components in the pathway between the clinical phenotype and the genotype, and emerge as powerful tools for gene discovery. Recent research highlights the value of compositional approaches in brain imaging studies, particularly in AD, where structural alterations emerge years before symptoms onset and changes in one region must be interpreted in the context of changes occurring across interconnected regions. The current study aims to identify relative brain volumetric variations in cortical and subcortical regions that (i) are associated with different AD stages along the disease continuum and (ii) vary by AD genetic risk. Using compositional data analysis (CoDA) we identified the relative variation of brain regions’ volumes, captured in compositional brain scores, associated with specific disease stages along the disease continuum. The genetic profile for AD shaped the way that these structures were interdependently interacting all along the continuum. CoDA provides an accurate analysis of brain imaging data accounting for the common overlooked compositional nature of the brain.