Comparative Features of MIA GMDH and Deep Feed-Forward Neural Networks
Olha H. Moroz, Volodymyr Semenovych Stepashko · Cybernetics and computer engineering · 2021
Introduction.Deep neural networks are effective tools for solving actual tasks such as data mining, modeling, forecasting, pattern recognition, clustering, classification etc.They differ with respect to the architecture design, learning methods and so on.Most simple and widely used are deep feed-forward supervised NNs.The purpose of the paper is to compare briefly main features of the deep feed-forward deterministic supervised networks with the Multilayered Iterative Algorithm of GMDH (MIA GMDH) and to formulate main ideas of constructing a new class of hybrid deep networks based on the MIA neural network.Methods.Most usable deep feed-forward supervised neural networks have been studied: multilayered perceptron, convolutional NN and some its modifications, polynomial neural networks, genetic polynomial neural network etc.Results.There was carried out a comparative analysis of main features of the MIA GMDH neural network with the characteristics of other deep deterministic supervised neural networks.The most promising approaches are identified to improve the performance of this network, particularly by hybridization with methods of computational intelligence.The main idea of building a new class of hybrid deep networks based on MIA GMDH is formulated.