Séminaire

Machine learning models require to train hundreds or thousands of variables. In addition, their performance strongly rely on the careful choice of some hyperparameters. The latter can correspond to regularization parameters, some structure on the learned parameters (e.g., groups or hierarchical trees) or both. In order to address their automated learning, various methods have been devised. For instance, when one has to set one or a few scalar hyperparameters, a popular strategy is to perform cross-validation on a coarse grid. However, such approach quickly becomes unfeasible when facing a large number of hyperparameters due to the curse of dimensionality. In this talk, I will present a generic bilevel optimization framework for tuning the hyperparameters. In addition, a particular attention will be devoted to the automated learning of group structure in the data.