Generative model of autoencoders self-learning on images represented by count samples

The paper substantiates the concept of autoencoders focused on automatic generation of compressed images. We propose a solution to the problem of synthesizing such autoencoders in the context of machine learning methods, understood here as learning based on the input images themselves (in the bootstrap spirit). For these purposes, a special representation of images has been developed using samples of counts of a controlled size (sampling representations). Based on the specifics of this representation, a generative model of autoencoders is formalized, which is then specified to a probabilistic parametric sampling model in the form of a mixture of components. Based on the concept of receptive fields, a reduction of the general model of a mixture of components to a grid model of finite components of an exponential family is discussed. This allows the synthesis of computationally realistic coding algorithms.