Head of the team : Franck Iutzeler. Permanent members.
The Statistics & Optimization team works on statistical estimation and optimization methods, two closely related fields, especially in the context of machine learning. The team's work covers a wide range of topics, including the following major themes :
Statistical Estimation & Tests : parametric and non-parametric estimation, statistical learning, independence and goodness-of-fit test, minimax optimality, differential privacy, conformal inference
Post-selection Inference : correction for model or variable selection effects, construction of valid confidence intervals, post hoc bounds on false positive rates
Global Sensitivity Analysis : Sobol'–Hoeffding decompositions and generalizations, Sobol', Shapley, and kernel indices, accounting for dependent entries
Gaussian Processes : asymptotic guarantees, high dimension, data/physics hybridization, experimental design, non-Euclidean inputs
Optimal Transport : estimation of the Wasserstein distance, convergence rates and limit theorems, distributionally robust optimization and applications in learning
Bias & Fairness in Learning : fairness constraint formulation via sensitivity analysis or regularization, counterfactual models for interpretability, legislative implications and standardization
Applications to Health & Biomedical Data : modeling of longitudinal data, prediction of rare events, digital twins, health economics, multi-view omics data, structured biological data
Inverse Problems : deconvolution, instrumental variable regression, applications in medical imaging, fMRI, Electrical Impedance Tomography, seismology
Continuous Optimization : convex analysis, inertial algorithms, discrete/continuous time analysis, non-smooth optimization, convergence rates, conditioning
Stochastic Optimization : noisy and non-differentiable EM methods, variance reduction, distributed and federated versions, limit distributions on non-convex functions
Optimization for Neural Networks : landscape of optimized functions in deep networks, quantization and Straight-Through Estimator, theoretical guarantees around automatic differentiation
Interactions :
Many members of the team interact with other IMT teams, and more broadly with the Toulouse research ecosystem, through both occasional and more structured collaborations, in particular via:
The team also maintains numerous collaborations with industrial partners (EDF, Onera) and public institutions (INSERM, Toulouse University Hospital). Additionally, the team is committed to reproducibility and actively contributes to the development of open-source packages.
Team life :
The Statistics and Optimization seminar takes place on Tuesdays at 11:15 am, usually in the Katherine Johnson room.
The Toulouse Multidisciplinary Optimization Seminar (SPOT) takes place on the first Monday of each month at 2:00 PM, usually in the Thesis Room (C002) at ENSEEIHT (N7), 2 rue Charles Camichel, 31000 Toulouse.