Study of Methods for Generating Pseudo-Examples for Neural Network Training

This paper investigates methods for generating pseudo-examples for neural network training within the framework of the developed Alternate Incremental Learning Algorithm. This algorithm alternates between using Extreme Learning Machine for adjusting the weights of only the output layer during the operational state and using backpropagation for adjusting all the weights during the sleep state. Previously, during the sleep state, the neural network was trained on pseudo-examples generated by adding noise to the features and using the previous version of the network to obtain the target variables. This study explores various methods for generating pseudo-examples, describing, analyzing, and evaluating their effectiveness within the Alternate Incremental Learning Algorithm. The experimental results demonstrate the impact of different pseudo-example generation methods on the training quality and performance of the neural network. The proposed methods and obtained results can contribute to the improvement of incremental learning algorithms for neural networks and their adaptability to new tasks.