One of the most important factors for improving the resolution of optical systems is to compensate for the aberrations (distortions) of the wave front. As a rule, whether special measuring devices (wavefront sensors) are used for such compensation or adaptive mirrors that perform iterative correction of the wavefront. However, often (for reasons of compactness or weight reduction), it is not possible to use the special equipment for measuring aberrations. To obtain certain information on the wave front, one can use the measured point spread function (PSF) or the intensity pattern in the focal plane. Methods of processing two PSFs (focal and nonfocal) with the help of neural networks are known. In this paper, we investigate the possibility of recognizing the wave front from a single intensity pattern in the focal plane. The technology of deep machine learning - convolutional neural network is chosen as the way for implementation. The idea of this technology lies in the alternation of convolutional and subsampling layers, for the purpose of efficient image recognition. Such approach will allow to optimize the process of compensation of optical system aberrations and to reduce the amount of required input data.