Séminaire Image : « Anomaly Detection with Pre-Trained Representations », Jeremie Pantin

Nous aurons le plaisir d’écouter Jérémie Patin, MCF CDD au GREYC.
Il donnera un séminaire IMAGE le jeudi 2 octobre 2025 à 14h en salle de séminaire F-200.

Titre : « Anomaly Detection with Pre-Trained Representations »

Résumé : 
In Anomaly Detection, relatively few studies have examined how pre-trained models can be used without fine-tuning or weight adaptation, particularly in multimodal scenarios. We first introduce a lightweight autoencoder approach for identifying textual anomalies, considering representation from encoder-based large language models. We then extend this perspective to a multimodal setting, focusing on brain tumor analysis with pre-trained text/vision models. Finally, we conclude with an xAI problem statement: Can we consider predicted anomalies as post-hoc explainers ?