Аннотация:In many problems of processing experimental data, lack of sufficient amount of experimental patterns with due representativity leads to poor performance of machine learning methods used to implement the necessary processing. Obtaining additional experimental data is often too expensive or even impossible. This study is devoted to development of the approach based on generating additional patterns by variational autoencoders, which are possibly capable to capture the distribution of initial data and to use it in data generation. At the example of an inverse problem in spectroscopy, it is demonstrated that enhancing the training dataset with generated patterns may improve the quality of the solution of the studied problem.
