(eng): The well known "beauty defect" of probabilistic neural networks is the biologically unnatural complete interconnection of neurons. Despite of deep reasons of this undesirable property, it can be removed by a special subspace approach without leaving the exact framework of Bayesian decision-making. The structural optimization based on EM algorithm is controlled by an infor- mation criterion. In the present paper the method has been applied to recognize unconstrained handwritten numerals.