Vol 25, No 3 (2016)

Over the last decade, the deep neural networks are a hot topic in machine learning. It is breakthrough technology in processing images, video, speech, text and audio. Deep neural network permits us to overcome some limitations of a shallow neural network due to its deep architecture. In this paper we investigate the nature of unsupervised learning in restricted Boltzmann machine. We have proved that maximization of the log-likelihood input data distribution of restricted Boltzmann machine is equivalent to minimizing the cross-entropy and to special case of minimizing the mean squared error. Thus the nature of unsupervised learning is invariant to different training criteria. As a result we propose a new technique called “REBA” for the unsupervised training of deep neural networks. In contrast to Hinton’s conventional approach to the learning of restricted Boltzmann machine, which is based on linear nature of training rule, the proposed technique is founded on nonlinear training rule. We have shown that the classical equations for RBM learning are a special case of the proposed technique. As a result the proposed approach is more universal in contrast to the traditional energy-based model. We demonstrate the performance of the REBA technique using wellknown benchmark problem. The main contribution of this paper is a novel view and new understanding of an unsupervised learning in deep neural networks.

Hybrid two dimensional (2D) wavelength-hopping time-spreading coding techniques have been focused now days for Optical Code Division Multiple Access (OCDMA) systems to increase the capacity of system. In this paper, design and comparative analysis of five different 2D coding techniques has been performed. These codes use Synchronized Quadratic Congruence sequences (SQC), Prime Code sequences (PC), Synchronized Prime Sequences (SPS) and One-Coincidence Frequency Hopping Code (OCFHC) for wavelength hopping and (n, w, λ_a, λ_c) one dimensional optical orthogonal codes for time spreading respectively along with Extended Reed Solomon (E-RS) codes. It has been observed that SQC/OOC code inspite of having larger value of cross correlation than other codes under consideration, out performs due to its ability to support a larger code weight. The comparison on the basis of BER shows that irrespective of the increase in number of hits due to higher code weight, SQC/OOC provides better code performance and hence supports large number of users in the system. The results are reported on the basis of maximum auto- and cross correlation function, cardinality and bit error rate (BER).

The characteristics of SpecTIR and AVIRIS 16-bit hyperspectral images are analyzed. The requirements for compression of such images are formulated. The aspects of using the hierarchical compression algorithm in hyperspectral images storage are studied. Spectral component approximation algorithms are considered that allow both an increased compression ratio and retrieval of particular components. Interpolation algorithms are considered and a rank interpolator is offered for hyperspectral images compression. Real 16-bit hyperspectral images are used in computational experiments to investigate the efficiency of the proposed algorithms. The best parameters of these algorithms are found experimentally and general recommendations on how to tune the proposed hierarchical compression algorithm to suit hyperspectral images storage problems are given.

The use of special digital image processing for removal of the blur which results quite often from a mutual relative motion of the optical sensor and object during the exposure is considered. The problem of blur removal from a digital image belongs to the class of incorrect problems, which means that it is impossible to obtain an exact solution resistant to small perturbations in the input data. Therefore, special methods should be invoked. The use of the Tikhonov regularization methods for solving the problem of blur removal from a digital image is demonstrated.

Problems of neural network forecasting system, invariant to type of energy consumption schedule are solved. Minimum length input vector structure is explained; neural network ensemble structures are determined; selection of the most effective neural network types in the ensemble is held. Original three-level structure of neural network ensemble is developed. Its high forecasting capability makes network perspective for solving information statistical analysis problems.