PhD Researcher in Topological Data Analysis for AI Anomaly Detection

This PhD position at INRIA’s Saclay Centre focuses on advancing topological data analysis (TDA) to evaluate and detect anomalies in complex deep learning models, particularly transformer-based architectures. The project, part of the national AI evaluation program (EVALIA), combines mathematical rigor with practical applications to study abnormal behaviors like hallucinations and adversarial attacks. The successful candidate will collaborate with the DataShape team (INRIA Saclay) and the LMO (Laboratoire de Mathématiques d’Orsay) to explore two research axes: analyzing activation structures in neural networks and dynamically tracking data distributions in latent spaces. The work will culminate in open-source tools integrated into the Gudhi library, a key resource in TDA.

• Conduct research on topological data analysis to detect and characterize anomalies in deep learning models, particularly transformer-based architectures.
• Develop mathematical tools with theoretical guarantees for evaluating AI model behaviors like hallucinations and adversarial attacks.
• Analyze activation structures in neural networks (e.g., attention matrices) using TDA and statistical methods.
• Track dynamic changes in data distributions across latent spaces over time.
• Implement and publish open-source tools for TDA, integrating them into the Gudhi library.
• Collaborate with the LMO (Laboratoire de Mathématiques d’Orsay) and DataShape team on interdisciplinary research.
• Present findings at academic conferences and workshops.
• Contribute to the broader EVALIA program for AI evaluation.