In this study the authors investigated the connections between the training processes of unsupervised neural network models with self-encoding and regeneration and the information structure in the representations created by such models. We propose theoretical arguments leading to conclusions, confirmed by previously published experimental results that unsupervised representations obtained under certain constraints in training compliant with Bayesian inference principle, favor configurations with better categorization of hidden concepts in the observable data. The results provide an important connection between training of unsupervised machine learning models and the structure of representations created by them and can be used in developing new methods and approaches in self-learning as well as provide insights into common principles underlying the emergence of intelligence in machine and biologic systems.