mardi 4 novembre 2025, de 14h à 15h30.
Dans le cadre des webinaires organisés par le CATI IMOTEP, Diego MARCOS (INRIA Montpellier) nous fera la présentation suivante (suivie de questions) :
"DeepMaxent: Improving Maxent Species Distribution Models with Neural Networks"
Les inscriptions se font via le lien suivant :
https://sondages.inrae.fr/index.php/478722?lang=fr
Résumé :
"Citizen science is rapidly expanding biodiversity data, but presence-only records come with strong sampling biases and lack absences, limiting their use in species distribution models (SDMs). Maxent, a widely used method based on the maximum entropy principle, relies on hand-crafted features and often struggles under biased data.
We introduce DeepMaxent, a neural network-based SDM that automatically learns complex, shared feature representations across species while retaining the maximum entropy framework. The model uses a normalized Poisson loss to estimate presence probabilities, enabling efficient learning from presence-only data.
In tests across several datasets, DeepMaxent consistently outperforms Maxent and other leading SDMs. It is particularly effective under uneven sampling, showing improved accuracy and robustness. The approach also scales well to large, multi-species datasets."