A Flexible Algorithm for Developing, Evaluating, and Visualizing Neural Network Models for Classification Tasks

This paper considers an approach to the development, evaluation, and continued training of neural network models for two-dimensional classification tasks using visual analysis of the results. The developed software system enables dataset generation, neural network configuration, model training, intermediate-state saving, calculation of numerical metrics, and visualization of complete classification regions. This makes it possible to evaluate models not only according to their loss values but also according to the shape of the boundaries between classes and the behavior of the network in individual regions of the feature space. Neural networks with different architectures and activation functions were compared experimentally. The obtained results showed that the lowest loss value does not always correspond to the most appropriate classification regions. Therefore, visual representation of the results should be used together with numerical metrics. The continued training of a model after the introduction of a new class was also investigated, with particular attention paid to the effect of the ratio between previously learned and newly introduced data on the preservation of existing classification regions. In the conducted experiment, the best result was obtained with a 2:1 ratio of old to new data. The proposed approach can be used to analyze how network architecture, model parameters, and the composition of the training dataset affect classifier behavior.