COMPARISON OF THE EFFICIENCY OF SOLUTION OF CLASSIFICATION TASK BY METHODS OF LINEAR DISCRIMINANT ANALYSIS AND ARTIFICIAL NEURAL NETWORKS

We compare two methods of classification: linear discriminant analysis and artificial feedforward neural networks. The comparison is carried out on the example of the task of classifying cyanobacteria in a series of self-fluorescence spectra of the individual cells. Methods of visualization and analysis of classification results, including new methods based on the entropy of classification results, have been demonstrated. The correlation between the results of linear discriminant analysis and classification of neural networks is determined. The key moments are shown, on the basis of which it is possible to reveal an incorrectly trained model of the neural network, and also a new method for analyzing the quality of the generalization is proposed. The methods of comparing the results of linear discriminant analysis and neural networks presented in the paper are applicable for comparing neural networks with other machine learning algorithms and statistical classifiers. The proposed method allows analyzing the learning process of a neural network in order to quickly identify critical deviations in its operation and possible illogical behavior in the future.

Authors: T. R. Zhangirov, А. S. Perkov, S. A. Ivanova, A. A. Liss, N. Yu. Grigoryeva, L. V. Chistyakova

Direction: Informatics, Computer Technologies And Control

Keywords: Classification, linear discriminant analysis, neural networks, statistical analysis, gradient descent


View full article