Pierre Fihey Adapting Anomaly Detection to Textual Representations
- Date: 9 novembre 2026 à 13h00
- Salle: 65-66 304
Textual anomaly detection usually relies on pre-trained language models by applying general-purpose anomaly detection methods to frozen representations. While this approach outperforms standalone models, recent benchmarks report highly heterogeneous results across detector–encoder pairs. Building on the intuition that these disparities stem from the highly anisotropic geometry of textual representations, which challenges many detectors, we investigate how to adapt this pipeline accordingly. In this talk, I will present our two recent publications, which explore complementary directions. The first one focuses on reshaping the representation space to make it more suitable for existing detectors. The second introduces a detector that directly captures the structure of the representation space, using data depth functions, that additionally provide robustness to contamination