Public Thesis Defense of Jérémie Bogaert - ICTEAM

Explanation Sensitivity to the Training Randomness in Text Classification Transformers by Jérémie Bogaert.

Le 16 janvier 2026 à 16h30 - Auditoire BARB92, place Sainte-Barbe à 1348 Louvain-la-Neuve.

Large language models (LLMs) that rely on the transformer architecture, such as BERT and GPT, have shown impressive performances in a variety of tasks and are now a cornerstone in natural language processing (NLP). Yet, understanding and reflecting their decision process has been recurrently shown to be difficult and is a major concern in many contexts in which they are used, especially when their decisions have important implications. Various supposedly desirable criteria for model explainability and methods offering different trade-offs have been introduced in the literature. A recent example used in this work is the Layer-wise Relevance Propagation (LRP) that produces explanations in the form of attention maps, assumed to be easily understandable by a human reader.

While explaining one instance of a classification transformer is already an important  challenge, these models are typically trained with stochastic optimization methods that vary depending on multiple sources of randomness such as parameter initialization, data shuffling, and dropout. It follows that even if two models trained on the same data with different randomness often achieve similar performances, nothing prevents them to provide different explanations of their predictions. Since there is no way to know a priori the impact that a certain randomness will lead to, it raises fundamental questions about the stability of classification transformers' explanations, which this work investigates. 

Jury members :

Prof.François-Xavier Standaert (UCLouvain), co-supervisor
Prof.Antonin Descampe (UCLouvain), co-supervisor
Prof.Laurent Francis (UCLouvain), chairperson
Prof.Cédrick Fairon (UCLouvain), secretary
Prof.Benoît Frénay (UNamur)
Dr.Loic Masure (CNRS)

Pay attention: the public defense of Jérémie BOGAERT will also take place in the form of a videoconference