Employing different loss functions for the classification of images via supervised learning (Q2440581)

In this paper, the support vector machines approach is applied to an image classification task. To this end, the authors reformulate the general Tikhonov regularization problem, to which the supervised learning problem gives rise, as a convex (not necessarily differentiable) optimization problem. They construct a conjugate dual to it, prove under suitable qualification conditions the existence of strong duality and express the optimal solutions of the primal problem via the ones of the dual. This has as a consequence the formulation of the decision function to be learned by means of the optimal solutions of the dual. Hence for the specific learning task one only has to numerically solve the dual problem, which, different to the primal one, can be mainly formulated as a convex differentiable optimization problem.