Anna van Elst

Since October 2024, I am a second year PhD student at Télécom Paris, Institut Polytechnique de Paris under the supervision of Prof. Stephan Clémençon and Igor Colin. My research is supported by the PEPR IA Foundry grant. I am part of the S2A team and a member of the LTCI lab. Prior to this, I completed an MSc in Informatics at TUM and obtained a Diplôme d’Ingénieur from Télécom Paris.

My research focuses on robust and decentralized machine learning, at the intersection of robust statistics, distributed optimization, and graph theory (see my Google Scholar). Specifically, I develop decentralized estimation and optimization methods with provable convergence and robustness guarantees. Unlike federated learning, decentralized approaches eliminate the need for a central server, avoiding communication bottlenecks and single points of failure. Most existing decentralized machine learning methods rely on gossip protocols and mean-based aggregation, which are highly sensitive to data contamination. To address this, I develop decentralized estimation methods for robust location parameters (e.g., medians and trimmed means) which offer higher breakdown points. I am also interested in quantile and rank-based estimators to detect extreme values and potential outliers in decentralized AI systems. Prior to this, I worked with Prof. Debarghya Ghoshdastidar on PAC-Bayesian generalization bounds for contrastive learning (SimCLR framework).