Том 26, № 4 (2017)

The paper is focused on the idea to demonstrate the advantages of deep learning approaches over ordinary shallow neural network on their comparative applications to image classifying from such popular benchmark databases as FERET and MNIST. An autoassociative neural network is used as a standalone program realized the nonlinear principal component analysis for prior extracting the most informative features of input data for neural networks to be compared further as classifiers. A special study of the optimal choice of activation function and the normalization transformation of input data allows to improve efficiency of the autoassociative program. One more study devoted to denoising properties of this program demonstrates its high efficiency even on noisy data. Three types of neural networks are compared: feed-forward neural net with one hidden layer, deep network with several hidden layers and deep belief network with several pretraining layers realized restricted Boltzmann machine. The number of hidden layer and the number of hidden neurons in them were chosen by cross-validation procedure to keep balance between number of layers and hidden neurons and classification efficiency. Results of our comparative study demonstrate the undoubted advantage of deep networks, as well as denoising power of autoencoders. In our work we use both multiprocessor graphic card and cloud services to speed up our calculations. The paper is oriented to specialists in concrete fields of scientific or experimental applications, who have already some knowledge about artificial neural networks, probability theory and numerical methods.

The problem of neural network model synthesis using the fractal analysis is addressed in the paper. The set of indicators characterizing the data sample properties from the unified position based on the fractal analysis principles is proposed. The methods of sample fractal dimension determining are proposed. These methods and indicators can be used to solve the problem of data dimensionality reduction. The set of indicators characterizing properties of neural network model is proposed. These indicators can be used in the process of neural network synthesis for the weights set contrasting, also as to find and remove the non-informative features and non-informative connections from the neural network model. The developed indicators and methods are implemented in software and studied at practical problem solving.

In this paper diffraction of a plane monochromatic TE-wave on an ideal homogeneous dielectric cylinder with several resonant wavelength scale radii is analyzed. Two subsequent near-surface maxima of intensity (two focuses) generated at the cylinder output were found on the optical axis. The first subwavelength focus is formed by one of the whispering gallery mode lobes. Its intensity is 50 times the incident light intensity and its full width at the half maximum of the intensity is equal to 0.155 of the incident wavelength. The second focus is two times less in intensity. Its focal spot known as a photonic nanojet is stretched toward the optical axis. The second focus is formed at a distance about the wavelength from the cylinder surface. Its width is equal to 0.44 of the wavelength and its length is two wavelengths. The abilities of light focusing by a two-layered cylinder and influence of materials absorption on the light focusing are also examined by numerical simulation.

The results of experimental research of working hybrid optical-digital image recognition system based on features extracted from Fourier spectra are presented. To form the features an optical Fourier processor with liquid-crystal input device is used. Recognition is carried out with computer (PC) using a neural network algorithm. The possibility of applying the developed system for recognition of the terrain images obtained by aerial photography is shown.