Nearest Class Vector Classification for Large-Scale Learning Problems

In this paper, we describe a method for combined metric learning and classification, that is based on logistic discrimination for the determination of a low-dimensional feature space of increased discrimination power. An iterating optimization process is applied to this end, where the probability of correct classification rate is increased at each optimization step. Extensions of the method that allow richer class representations and non-linear feature space determination and classification are also described. The described optimization schemes are solved by following (stochastic or mini-batch) gradient descent optimization, which is well suited for large-scale learning problems.