Deep Neural Networks and Hybrid GMDH-Neuro-fuzzy Networks in Big Data Analysis

Deep neural networks (DNN) and their generalizations—hybrid DNN are considered which represent the efficient tools for BD analysis. The properties and drawbacks of deep learning are considered and analyzed. Encoders—decoders and restricted Boltzmann machines are described and their applications for Deep learning implementation are presented. Methods of regularization of DL: penalty functions, Dropout and Bagging are presented. New class of deep learning networks are suggested and presented. so-called GMDH-neo-fuzzy networks representing a combination of self-organization method GMDH and fuzzy neural networks. Due to principle of self-organization and small number of tuning parameters GMDH enables to simplify and accelerate the training of DN. Several variants of this class hybrid networks are considered and algorithms of their structure synthesis based on GMDH are suggested and analyzed. The application of GMDH enables to reduce dimensionality of training DN and accelerate the convergence of training and by this solve some problems of Big Data Analysis. Experimental investigations of hybrid GMDH-neo-fuzzy networks are carried out and their results are presented and analyzed.