The Belgium-Brazil-Luxembourg Graduate Seminar on Applied Mathematics is a space dedicated to scientific and business collaboration between Belgium, Brazil and Luxembourg. This initiative is sponsored by Brascam.

Please contact me if you are interested in submiting a talk.

Join the meeting.

Subscribe to our mailing list by clicking here.

Next Edition: 25th of June, 2025

  • Theme: Artificial Intelligence and Machine Learning
  • Brussels Time: from 16:00 to 17:00
  • Brasilia Time: from 11:00 to 12:00
  • Networking: anyone interested in networking is welcomed to arrive 10 minutes before the start and stay after the end of the talks in the meeting room.

Karolayne Dessabato - Universidade Federal do Rio de Janeiro (UFRJ)

  • Title: Explainable Machine Learning in Music Emotion Recognition Using Shapley Values
  • Abstract: Music Emotion Recognition (MER), an area within Musical Information Retrieval (MIR), studies the emotions evoked in listeners by music. We address MER as a regression task, with the objective of predicting the emotional content of music (encoded in arousal and valence) from acoustic features extracted from the waveform. We apply an interpretable machine learning technique, investigating the role of these features in predicting the target variables. Initially, a random forest model is trained on the DEAM dataset (MediaEval Database for Emotional Analysis of Music). Then, we use the concept of Shapley values to interpret the role of each variable in the predictions made by this model. Finally, we extract the most significant features from the DEAM dataset to predict arousal and valence, thus enhancing the interpretability of the model employed.

Luís Espírito Santo - Universidade de Coimbra (UC) and Vrije Universiteit Brussel (VUB)

  • Title: Introduction to Creativity and Steps Towards a Formal Creativity Theory
  • Abstract: The growing influence of generative AI has reignited interest in how artificial systems engage with creative tasks, often centering discussions around large language models and their impact on human creativity. Yet, the field of Computational Creativity—active since the 1990s—offers alternative perspectives that can help reframe these debates. In this talk, we revisit some of these foundational ideas and explore a recent proposal that tries to pave the way for a Formal Creativity Theory as a new direction that draws on Formal Learning Theory to better understand creativity.

Past Editions

28th of May, 2025

Josué Knorst - Universidade Federal do Rio Grande do Sul (UFRGS)

  • Title: Ergodicity, mathematically put
  • Abstract: In mathematics, ergodicity describes a property of a measure with respect to a space transformation. This property is not difficult to attain; indeed, numerous examples can be found in the periodic orbits of a dynamical system. After Birkhoff’s theorem, this concept was shown to be equivalent to various other perspectives on a system, ranging from recurrence properties to the convergence of time averages. We will illustrate these features through concrete examples.

Arne Vanhoyweghen - Vrije Universiteit Brussel (VUB)

  • Title: Ergodicity Economics: A Behavioral and Mathematical Framework
  • Abstract: Ergodicity Economics addresses how averages across time and across a population relate in practical settings. Many systems in economics, finance, and even personal decision-making are often assumed to be ergodic — meaning that an individual’s time-averaged outcomes are representative of the average outcomes across a whole population. However, this assumption usually doesn’t hold in real-world scenarios. Exploring human decision making from the lens of time averages as opposed to ensemble averages, sheds a new light on widely accepted cognitive peculiarities in human decision making such as loss aversion and status quo bias. In this talk we will discuss how these behaviors follow as a natural consequence of optimizing one’s trajectory over time, and how this insight can help us shape policy design.

30th of April, 2025

Matheus Oliveira - Universidade Federal do Rio de Janeiro (UFRJ)

  • Title: Streptomyces: Next Gen Bioinoculants Against Plant Diseases
  • Abstract: Brazilian agriculture faces significant losses due to phytopathogenic fungi, terribly affecting crops like sugarcane. Thus, research related to biological control of phytopathogens has been gained prominence in recent decades. Actinobacteria are notable for their metabolic versatility (production of plant hormones, siderophores, enzymes and antibiotics) in biological control. Therefore, offering a wide range of applications and physical-chemical effects, presenting themselves as a promising solution for bioinoculants development.

Vincent Crabbe - Fonds voor Wetenschappelijk Onderzoek - Vlaanderen (FWO) and Vrije Universiteit Brussel (VUB)

  • Title: Engineering Bacteria to Sense Heavy Metal Pollution
  • Abstract: Mining is a key economic driver in many Latin American countries, including Peru and Brazil. However, it also leads to heavy metal pollution, which poses risks to both the environment and human health. Unfortunately, regular monitoring is often lacking because traditional detection methods are expensive and require specialised expertise. In this presentation, Vincent will explain how engineered bacteria can serve as innovative heavy metal sensors and present findings on an arsenic biosensor developed at VUB.

26th of March, 2025

Marlon Moresco - Universidade Federal do Rio Grande do Sul (UFRGS)

  • Title: Constructing Elicitable Risk Measures
  • Abstract: We provide a constructive way of defining new elicitable risk measures that are characterised by a multiplicative scoring function. We show that depending on the choice of the scoring function’s components, theresulting risk measure possesses properties such as monotonicity, translation invariance, convexity, and positive homogeneity. Our framework encompasses the majority of well-known elicitable risk measures includingall elicitable convex and coherent risk measures. Our setting moreover allows to construct novel elicitablerisk measures that are, for example, convex but not coherent. Furthermore, we discuss how higher-orderelicitability, such as jointly eliciting the mean and variance or different quantile levels, fall within our setting.
  • Slides

Morten Wilke - Vrije Universiteit Brussel (VUB)

  • Title: Optimal Payoffs under Smooth Ambiguity
  • Abstract: We study optimal payoff choice for an investor in a one-period model under smooth ambiguity preferences, also called KMM preferences as proposed by Klibanoff et al. (2005). In contrast to the existing literature on optimal asset allocation for a KMM investor in a one-period model, we also allow payoffs that are non-linear in the market asset. Our contribution is fourfold. First, we characterize and derive the optimal payoff under KMM preferences. Second, we demonstrate that a KMM investor solves an equivalent problem to an investor under classical subjective expected utility (CSEU) with adjusted second-order probabilities. Third, we show that a KMM investor with exponential ambiguity attitude implicitly maximizes CSEU utility under the ‘worst-case’ second-order probabilities determined by his ambiguity aversion. Fourth, we reveal that optimal payoffs under ambiguity are not necessarily monotonically increasing in the market asset, which we illustrate using a log-normal market asset under drift and volatility uncertainty.
  • Slides

26th of February, 2025

Althayr Santos - Universidade Federal do Rio de Janeiro (UFRJ)

  • Title: Updating Portuguese BERT: A Modern take on Training Encoder-type Large Language Models
  • Abstract: This talk presents a comprehensive approach to building and openly distributing a state-of-the-art BERT model for Portuguese using publicly available data. I will detail the proposed updates to the original architecture and discuss the practical challenges of distributed training on a Nvidia DGX cluster. We expect the resulting model to improve upon the original BERT across virtually all tasks while maintaining its compact size. The final model, along with all supporting data and code, will be open-sourced with a permissive license. It can serve as a foundation for various NLP tasks such as text and token classification (sentiment analysis, PII recognition) and semantic embedding generation (for search and recommendation systems, RAG-based LLM applications, and AI agents).
  • Slides

Marthe Ballon - Vrije Universiteit Brussel (VUB)

  • Title: The Geometry of Thought: Where LLMs Go to Work Their Math-gic
  • Abstract: The text generation process of Large Language Models (LLMs) relies partly on converting tokens into high-dimensional vectors - called embeddings - and back. These embeddings constitute an ultra-high dimensional vector space. Recently, the structure of the embedding space has become a subject of study in the literature on interpretability and control of LLMs. For example, recent research shows that the token space of the LLaMa-2 model contains linear representations of high-level concepts such as man->woman, which allows one to exploit linear algebra operations to measure and manipulate the LLM’s behaviour (Park et al. 2024). Our approach starts from the idea that we can discover - and tune - how LLMs learn to reason about math by analysing a.o. the geometry of the embedding space. In this presentation, we will analyse the embedding space of an Olympiad-level math dataset, generated by the latest embedding model of OpenAI. Although this talk is more introductory, it will illustrate how much information can be extracted from the embedding space and how it can give us new insights into the reasoning abilities of LLMs in mathematics.
  • Slides

29th of January, 2025

Rodrigo Ribeiro - Fundação Getúlio Vargas (FGV)

  • Title: Belief Influence and Optimal Execution: A Mean Field Game Approach.
  • Abstract: In financial markets, large trading firms often influence the collective beliefs of market participants, creating conflicts of interest that regulatory efforts address through “Ethical Walls.” We propose a mean field game model to analyze how a major player optimally trades while incurring costs to influence the beliefs of minor agents. The framework leads to a Nash equilibrium characterized by coupled Hamilton-Jacobi-Bellman equations, which simplify into a system of ordinary differential equations. Numerical simulations illustrate the equilibrium dynamics, offering insights into regulatory and strategic considerations for large market participants.
  • Slides

Alesia Gerassimenko - Vrije Universiteit Brussel (VUB)

  • Title: Validating spatial dynamics for energy efficiency in the Belgian residential rent market.
  • Abstract: A dataset of 22,834 listed Belgian rental properties in Belgium is analysed to examine the relationship between energy efficiency and rental prices, using the four most commonly applied econometric models: Ordinary Least Squares (OLS) model, Generalized Additive Model (GAM), Spatial Durbin Model (SDM) and Geographically Weighted Regression (GWR) model are analysed. While this study underscores the importance of selecting the appropriate research strategy in general, it specifically highlights the importance of incorporating geographical information into the methodology of energy efficiency studies.
  • Slides