Machine Learning in Montpellier, Theory & Practice – Guilherme Dean Pelegrina (Mackenzie Presbyterian University, Brazil)
Many applications involve estimating a source of interest from mixed signals collected by multiple sensors. While many works focus on optimization task for source extraction, less attention has been given to interpreting how each sensor contributes to the result. In this talk, Guilherme presents a model-agnostic, game-theoretic approach based on Shapley values to quantify sensor contributions and interaction effects, with experiments on synthetic and real data. Eugenio Dias Ribeiro Neto (LIRMM): Title: An improved architecture for part-based animal re-identification through semantic segmentation distillation Abstract: Animal re-identification is essential for non-invasive wildlife monitoring but remains challenging due to limited data and high appearance variability. Eugenio will present PAW-ViT, a part-aware Vision Transformer that uses learnable part tokens specialized for anatomical regions. The method achieves state-of-the-art results on animal Re-ID benchmarks, especially under strong viewpoint changes.
