COMPARISON OF NEURAL NETWORK TRAINING METHODS IN THE CLASSIFICATION PROBLEM
Based on the analysis of a series of feedforward artificial neural networks, a method has been developed for determining the optimal neural network architecture for the task of classifying cyanobacterial strains according to the fluorescence spectra. The analysis of six gradient methods of training neural networks and their parameters was carried out, the optimal number of neurons in the hidden layer for neural networks trained by each of the methods was found, various methods of initializing the weights of neurons and methods for splitting the initial sample into training, test and control samples were evaluated. The choice of the optimal architecture was carried out on the basis of the classification results, namely, on the basis of the classification accuracy graphs and the classification error graphs. The research was conducted on the example of recognition of 16 classes, representing 16 strains of cyanobacteria. A number of shortcomings were identified in the method of testing feedforward neural networks and directions for additional researches of neural networks for classification in terms of extending the testing methodology of their internal logic were determined.
Authors: А. S. Perkov, T. R. Zhangirov, A. A. Liss, N. Yu. Grigoryeva , L. V. Chistyakova
Direction: Informatics, Computer Technologies And Control
Keywords: Neural networks, training methods, machine learning, classification problems, initialization methods
View full article