Research
Statistical methods for learning, inference and decisions with complex data.
My research develops and analyzes statistical learning methods: estimating difficult objects, learning from high-dimensional data, controlling generalization, combining decentralized information and remaining reliable under noise or adversarial behavior. It connects statistical estimation theory and inverse problems with contrastive learning, clustering and robustness.
High-dimensional clustering and consensus
CoHiRF — hierarchical consensus, stability and interpretability
CoHiRF builds repeated partitions on randomly selected feature subspaces and consolidates them through consensus. Rather than relying on one potentially unstable partition, it combines complementary local views into a hierarchical clustering representation. Medoids and the fusion hierarchy make successive merges inspectable. The method is designed for high-dimensional and large-scale data and can work with different underlying clustering procedures. Numerical studies investigate noise robustness, stability and scalability. Manuscript submitted in 2026.
VertCoHiRF — decentralized vertical federated clustering
Vertical federated learning involves agents holding different features or modalities for the same observations. VertCoHiRF combines locally generated partitions without centralizing raw features: agents exchange observation identifiers and cluster labels. It accommodates heterogeneous local methods and objectives and investigates Byzantine robustness, communication requirements and consensus formation. Manuscript submitted in 2026. A further extension for very large sample sizes has been developed separately and is not yet submitted.
Contrastive learning, conditional densities and causal inference
MCD — Marginal Contrastive Discrimination
MCD turns conditional density estimation into discrimination between the joint distribution and the product of its marginals. A contrastive classifier provides an estimate of a density ratio, from which the conditional density can be recovered. Our work studies the statistical theory linking classification risk to density-estimation error, as well as practical implementations based on mixing and shuffling observations. A central motivation is to handle high-dimensional covariates without directly estimating a high-dimensional joint density. Ongoing experiments assess predictive accuracy across architectures and data-generating models. An extension to multivariate target variables is under development. Methodological and theoretical manuscript in preparation.
Counterfactual densities and causal inference
The MCD framework also motivates the estimation of counterfactual outcome distributions, beyond average treatment effects. Estimating full distributions helps characterize heterogeneous effects, uncertainty and tail behavior under different interventions or policies. A first contribution with Aminata Ndiaye and Madalina Olteanu appeared at ESANN 2026; further methodological and theoretical developments are ongoing.
Neural network generalization and small-data learning
AdaCap — adaptive capacity control
AdaCap studies how to train neural networks on small tabular datasets while controlling overfitting. It combines permutation-based contrastive regularization with a Tikhonov-regularized output layer admitting a closed-form solution. The goal is to adjust effective model capacity to available information. The method was presented at ESANN 2026 and evaluated across a broad set of regression datasets.
MLR — Muddling Labels for Regularization
MLR uses perturbed or shuffled labels to inform regularization. Beyond the practical loss, the research addresses the relationship between network capacity, generalization and statistical regularization. A separate theoretical paper is being developed and should not be counted as an additional published AdaCap article.
Robust, adversarial and distributed learning
Learning under adversarial agents
A developing research direction studies the trade-offs between Byzantine resilience, computational efficiency and data or model ownership verification. Topics include robust aggregation in distributed learning, reproducibility of empirical robustness claims and statistical guarantees under adversarial behavior. A PhD co-supervision on these topics is being prepared; the project is not presented as an active enrolled doctorate.
Mathematical statistics, inverse problems and quantum information
Quantum tomography and nonparametric inference
My earlier research addresses statistical reconstruction of quantum states from indirect, noisy measurements, especially homodyne tomography. Contributions include adaptive Wigner-function and density-matrix estimation, nonparametric testing, quadratic functional estimation and minimax analysis. These problems combine ill-posed inverse operators, regularization and precise risk bounds. The results appeared in Annals of Statistics, Inverse Problems, Physical Review A and Electronic Journal of Statistics. This background in inverse problems and statistical inference also informs future questions in quantum information and computation; no unpublished quantum-computing result is claimed.