N. T. Abdullaev – Ph.D. (Eng.), Head of Department of Biomedical Engineering, Azerbaijan Technical University (Baku)

E-mail: a.namik46@mail.ru

K. Sh. Ismaylova – Ph.D. (Eng.), Associate Professor, Azerbaijan State University of Oil and Industry (Baku)

Many utilized artificial neural networks have a monologue structure. Such networks are well executed with small output data, but with increase in the input data the network complexity increases and their possibilities decrease. The aim of the work is using of artificial neural networks with modular architecture to improve the accuracy of diagnostic solutions of input parameters for electromyographic signals with similar values.

Due to the proposed modular architecture, the number of weight connections is less than in a fully connected multilayer perceptron. A modular network is designed to combine two different generalization approaches, known from connections and logical neural networks. It increases the possibilities of network generalization.

The proposed network system consists of a layer of input modules and an additional decision module. All subnets are MLP. Each input variable is connected to only one of the input modules. These connections are randomly selected, and the outputs of all input modules are connected to the decision network.

At the first stage, all the auxiliary networks in the input layer are trained, the test suite for each subnet is selected from the initial test suite. All input modules can be easily prepared in parallel, because they are all interdependent. At the second stage, the system of decision-making is taught. The training set for the decision module is built on the output of the input layer together with the original class number. To calculate the set, each input template is applied to the input layer, the resulting vector together with the desired output form a training pair for the decision module.

The experimental results have shown that the modular structure of the neural network is more effective, since for the multilayer perceptron structure the average values of the criterion for sensitivity were 86,19, and for specificity they were equal to 97,95. At the same time, these indicators for the modular structure were 90,91 (for sensitivity) and 99,61 (for specificity). The generalizing advantages for high-dimensional input vectors and the ease of parallel learning at the input level due to the independence of the modules provide the basis for the possibility of using the proposed structure in diagnosing diseases in myography.

1. Mauntkasl V. Organizuyushchij printsip funktsii mozga – elementarnyj modul' i raspredelennaya sistema / V kn. Edelmen Dzh., Mauntkasl V. Razumnyj mozg. M.: Mir. 1981.
2. Dorogov A.Yu., Alekseev A.A., Butorin D.A. Kontseptsiya strukturnogo modelirovaniya nejronnykh setej // Izvestiya VUZov. Ser. «Priborostroenie». 1998. T. 41. № 4. S. 36–41.
3. Kohonen T., Barna G., Chrisley R. Statistical pattern recognition with neural networks: benchmarking studies // Proc. IEEE International Conference on Neural Networks. San Diego. 1988. P. 61–67.
4. Mui L., Agarwal A., Gupta A., Wang P.Sh.-P. An adaptive modular neural network with application to unconstrained character recognition // International Journal of Pattern Recognition and Artificial Intelligence. 1994. V. 8. № 5. P. 1189–1204.
5. Bellotti R., Castellano M., De Marzo C., Satalino G. Signal/background classification in a cosmic ray space experiment by a modular neural system // Proc. of the SPIE – The International Society for Optical Engineering. 1995. V. 2492. № pt.2. P. 1153–61.
6. Boers E.J.W., Kuiper H. Biological metaphors and the design of modular artificial neural networks // Master’s thesis. Departments of Computer Science and Experimental and Theoretical Psychology at Leiden University, the Netherlands. 1992.
7. Chiang Ch.-Ch., Fu H.-Ch. Divide-and-conquer methodology for modular supervised neural network design // Proc. of the 1994 IEEE International Conference on Neural Networks. Part 1 (of 7). Orlando, FL, USA. 1994. P. 119–124.
8. Happel B.L.M., Murre J.M.J. Design and evolution of modular neural architectures // Neural Networks. V. 7. № 6/7. 1994. P. 985–1004.
9. Ishikawa M. Learning of modular structured networks // Artificial Intelligence. 1995. V. 75. № 1. P. 51–62.
10. Kim J., Ahn J., Kim Ch.S., Hwang H., Cho S. Multiple neural networks using the reduced input dimension // Proc. ICASSP – IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway, NJ, USA, 1994. V. 2. P. 601–604.
11. Kimoto T., Asakawa K., Yoda M., Takeoka M. Stock market prediction system with modular neural networks // 1990 International Joint Conference on Neural Networks – IJCNN 90. Piscataway, NJ, USA. 1990. P. 1–6.
12. Lu B.-L., Ito K., Kita H., Nishikawa Y. Parallel and modular multi-sieving neural network architecture for constructive learning // Proc. of the 4th International Conference on Artificial Neural Networks. Cambridge, UK. 1995. P. 92–97.
13. Schmidt A., Bandar Z. A modular neural network architecture with additional generalization abilities for large input vectors // Third International Conference on Artificial Neural Networks and Genetic Algorithms (ICANNGA-97). Norwich, England. 1997.
14. Tarkhov D.A. Nejrosetevye modeli i algoritmy. Spravochnik. M.: Radiotekhnika. 2014.