“Machine learning and optimization” working group meeting
 “Machine learning and optimization” working group meeting 6-Nov-2018 10:00 Age: 202 days The first session of the working group will be on November 6th, 2018, in the seminar room of CERMICS (2nd floor of Coriolis building, left wing, see here for instructions to reach the building). There will be two lectures: 10h-11h, Alexandre d’Aspremont, Restarting Frank-Wolfe Conditional Gradients (aka Frank-Wolfe algorithms) form a classical set of methods for constrained smooth convex minimization due to their simplicity, the absence of projection step, and competitive numerical performance. While the vanilla Frank-Wolfe algorithm only ensures a worst-case rate of O(1/\epsilon), various recent results have shown that for strongly convex functions, the method can be slightly modified to achieve linear convergence. However, this still leaves a huge gap between sublinear O(1/\epsilon) convergence and linear O(\log 1/\epsilon) convergence to reach an $\epsilon$-approximate solution. Here, we present a new variant of Conditional Gradients, that can dynamically adapt to the function’s geometric properties using restarts and thus smoothly interpolates between the sublinear and linear regimes. 11h15-12h15, Guillaume Obozinski, Structured prediction I will make an overview of structured prediction a.k.a. structured output learning. A number of problems in statistics and machine learning go beyond regression or multi-class classification, and require to make a more structured decision in the presence of uncertainty: some example include historically image segmentation, pose estimation, and a number of task in NLP such as part of speech tagging, word alignment of sentences from a source language to a target language in machine translation; it is still relevant today for alignment of knowledge bases, for applications in computational biology such as protein function prediction, or side-chain prediction, etc. These problems are very similar in structure to problems considered in operations research, except that the cost defining the objective, which is often a function of input data, is not known and must be learned. This type of formulation could be relevant for operations research problems in the presence of uncertainty when the distribution of the random phenomena to take into account is not known. Several frameworks and formulation have been proposed for structured prediction. I will present the main ones: graphical models and structured SVMs, and discuss some of the algorithms used for learning and prediction. Time permitting I will also talk about the “learn to search” paradigm and deep learning approaches. More information on bezout.univ-paris-est.fr/working-groups/learning-optimization/ <- Back to: Accueil
Dernière mise à jour : 04/06/2013