Bi-level Stochastic Gradient Methods for Large Scale Classification under Uncertainty

par Nicolas Couellan (Université de Toulouse)

vendredi 11 déc. 2015, 10:30 → 11:30 Europe/Paris

XLIM Salle X.203

In this talk, we will present bi-level stochastic gradient methods for classification with large scale support vector machines. We will first address the issue of large scale binary classification when data is subject to random perturbations. The proposed model integrates a learning framework that adjusts its robustness to noise during learning. The method avoids over-conservative situations that can be encountered with worst-case robust support vector machine formulations. The magnitude of the noise perturbations that should be incorporated in the model is controlled by optimizing a generalization error. No assumption on the distribution of noise is taken. Only rough estimates of perturbations bounds are required. The proposed bi-level algorithm performs very cheap stochastic sub-gradient moves and is therefore well suited to large datasets. We will present encouraging experimental results confirming that the technique outperforms robust second order cone programming formulations. Next, we will also show that the proposed method can be used to perform automatic selection of the penalty parameter in linear and non-linear support vector machines to avoid expensive k-fold cross validation computations.