This article is concerned with the acquisition of multimodal integration capacities by learning algorithms. Humans seem to perform statistically optimal fusion, and this ability may be gradually learned from experience. In order to stress the advantage of learning approaches in contrast to hand-coded models, we propose a generative-discriminative learning architecture that avoids simplifying assumptions on prior distributions and copes with realistic relationships between observations and underlying values. We base our investigation on a simple self-organized approach, for which we show statistical near-optimality properties by explicit comparison to an equivalent Bayesian model on a realistic artificial dataset.