Dolotov A. Self-learning spiking neural networks in data mining tasks

Research object - process of numeric data mining. Research target is the synthesis of data clustering methods on base of self-learning spiking neural network under a priori and current uncertainty - when clusters of data being processed overlap, are of complex form, or their amount varies with time, that ensure data processing speed increase as well as the synthesis of analog-digital architecture of self-learning spiking neural network. Methods of research: theory of computational intelligence - to define the research methodological foundations and context; theory of artificial neural networks - to analyze spiking neural network architecture, to improve its learning method and to modify multilayered spiking neural network; computational neuroscience - to analyze architecture and self-learning capability of spiking neural network and to synthesize its analog-digital architecture; fuzzy logic theory and cluster analysis - to synthesize hybrid fuzzy systems on base of spiking neural network; inductive modeling and design of experiments theory - to synthesize varying architectures of spiking neural network; automatic control theory - to synthesize analog-digital architecture of spiking neural network; simulation modeling - to confirm efficiency of the designed systems employment. Theoretical and practical results of the paper solves in total scientific task of data clustering under a priori and current uncertainty on base of self-learning spiking neural networks. Scientific novelty: 1) hybrid self-learning spiking neural network for fuzzy data clustering is proposed for the first time, that makes it possible to increase data processing speed when clusters to be detected overlap as compared to known conventional methods of fuzzy clustering; 2) fuzzy receptive neuron and, on its basis, architecture of input data fuzzification layer of self-learning spiking neural network is proposed for the first time, that in contrast to the known input layer of population coding makes it possible to increase data processing efficacy when there is a priori knowledge about task being solved; 3) analog-digital architecture of basic self-learning spiking neural network based on the Laplace transform is proposed for the first time, that makes it possible to describe biologically plausible neural networks functioning in terms of classical theory of automatic control; 4) self-learning multilayered spiking neural network is improved by removing adjustable lateral connections and decreasing number of learning methods down to one as opposite to the network proposed initially; 5) spiking neural network learning method is improved, it updates not only neuron-winner synaptic weights, but its neighbours ones also that in contrast to original method makes it possible to increase quality of learning of the proposed hybrid systems. Degree of implementation - the research results are used in the Shopping Center "Mars", Petrovske (act of 18.05.2010) and in the State Scientific Production Enterprise "Systemni Tekhnolohii", Dnipropetrovsk (act of 14.06.2010); scientific statements, conclusions, and recommendations contained in the thesis are used in courses "Neural network methods of computational intelligence" and "Data mining" which are taught to students of the specialty "Intelligent Decision Support Systems" of Kharkiv National University of Radio Electronics (act of 06.09.2010). The scope of use - in organizations that deal with problems of fuzzy data clustering intelligent systems development; various areas where it is necessary to group input data under uncertainty, particularly in video surveillance systems, for image segmentation, specific objects detecting on images, etc.; in the education process for preparing specialists in the area of intelligent information processing.