REDUCING THE IMPACT OF TRAINING DATA ERRORS ON TRAINING OF NEURAL NETWORKS IN MEDICAL EDUCATION

Tarasov V. S.

intelligence systems are actively being introduced into medical education. However, the reliability of AI systems depends on the quality of training data, which can be contaminated with various types of errors. These errors can lead to incorrect training results and have a negative impact on the quality of training and even on the health of patients. The main goal is to study the impact of various types of errors in training data on the training of neural networks used in medical education and to develop methods for eliminating their negative impact based on existing methods and algorithms. This article applies the following methods: classification of error types in training data, analysis of existing algorithms for reducing noise in training data and identification of their limitations with respect to intentional errors, modeling of neural network training and development and application of an improved EM (Expectation-Maximization) algorithm that takes into account the time it takes students to solve problems for a more accurate estimate of the noise distribution in labels. The experiments demonstrate a significant increase in the accuracy of neural network training when using the improved EM algorithm compared to the traditional approach.

Keywords:noise labels, training accuracy, problem solving time, adaptive optimization, feature vector, EM algorithm