PROBLEMS OF THE NEURAL NETWORKS OUTPUT DATA QUALITY ASSESSMENT

Today, artificial intelligence, particularly neural networks, is increasingly being used in software in a variety of industries, from mission-critical applications such as healthcare and the military to commerce and entertainment. One of the main stages of development and implementation of such software is the stage of quality control. To prevent fatal errors and to survive in a highly competitive environment, the software needs proper testing taking into account the peculiarities inherent in the data obtained as a result of the neural network. This article presents the relevance of using artificial intelligence systems in general and neural networks in particular and analyzes the main challenges that arise when assessing the quality of such networks. The author compares the properties of the output data of the artificial intelligence systems of the previous generation and the latest neural networks, highlights the key differences of the latter, such as the potential infinity of the input data sets and their relative unpredictability, the dependence of the results on the network training stage, and the subjective nature of the evaluation of such results. Based on the analysis, the author formulates a set of problems that can be solved using mathematical algorithms and methods. The main part of an article contains a general overview of existing solutions, with an emphasis on such algorithms and methods as calculating accuracy and loss, finding the F-score, interpretation methods and imitation modeling. As a result of the research, the author comes to the conclusion that, despite a sufficient number of existing solutions that can be used to solve the highlighted problems, they still have to be improved to increase the accuracy of neural network evaluation, as one hundred percent accuracy in evaluating data obtained as a result of the operation of neural networks has not yet been achieved.