S.V. Zimina

Lobachevsky State University of Nizhny Novgorod (Nizhny Novgorod, Russia)

Setting up artificial neural networks using iterative algorithms is accompanied by fluctuations in weight coefficients. When an artificial neural network solves the problem of allocating a useful signal against the background of interference, fluctuations in the weight vector lead to a deterioration of the useful signal allocated by the network and, in particular, losses in the output signal-to-noise ratio. The goal of the research is to perform a statistical analysis of an artificial neural network, that includes analysis of losses in the output signal-to-noise ratio associated with fluctuations in the weight coefficients of an artificial neural network.

We considered artificial neural networks that are configured using discrete gradient, fast recurrent algorithms with restrictions, and the Hebb algorithm. It is shown that fluctuations lead to losses in the output signal/noise ratio, the level of which depends on the type of algorithm under consideration and the speed of setting up an artificial neural network.

Taking into account the fluctuations of the weight vector in the analysis of the output signal-to-noise ratio allows us to correlate the permissible level of loss in the output signal-to-noise ratio and the speed of network configuration corresponding to this level when working with an artificial neural network.

Zimina S.V. Analysis of neural networks efficiency when accounting for weight coefficients jitter, leading to reduction in output signal/noise ratio. Neurocomputers. 2021. V. 23. № 2. Р. 15−25. DOI: https://doi.org/10.18127/j19998554-202102-02 (in Russian).

1. Bershad N.J., Eweda E., Bermudez J.C.M. Stochastic analysis of an adaptive line enhancer/canceler with a cyclostationary input. IEEE Transactions on Signal Processing. 2016. V. 64. № 1. P. 104-119. DOI: 10.1109/TSP.2015.2486745
2. Jeong J.J., Kim S.H., Koo G., Kim S.W. Mean – square deviation analysis of multiband – structured subband adaptive filter algorithm. IEEE Transactions on Signal Processing. 2016. V. 64. № 4. P. 985-994. DOI: 10.1109/TSP.2015.2498136
3. Eweda E., Bershad N.J. Stochastic analysis of the signed LMS algorithms for cyclostationary white gaussian inputs. IEEE Transac-tions on Signal Processing. 2017. V. 65. № 7. P. 1673-1684. DOI: 10.1109/TSP.2016.2646666
4. Li Z., Xia Y., Pei W., Wang K., Mandic D.P. An augmented nonlinear LMS for digital self – interference cancellation in full – duplex direct – conversion transceivers. IEEE Transactions on Signal Processing. 2018. V. 66. № 15. P. 4065-4078. DOI: 10.1109/TSP.2018.2846250
5. Khaykin S. Neyronnyye seti: polnyy kurs. 2-e izd., ispr. Per. s angl. M.: OOO «I.D. Viliams». 2006. 1104s. (in Russian).
6. Galushkin A.I. Teoriya neyronnykh setey. Kn.1: Ucheb. posobiye dlya vuzov. Obshchaya red. A.I. Galushkina. M.: IPRZhR. 2000. 416 s  (in Russian).
7. Tatuzov A.L. Neyronnyye seti v zadachakh radiolokatsii. Kn.28. M.: Radiotekhnika. 2009. 432 s (in Russian).
8. Zeng X., Yeung D.S. Sensitivity analysis of multilayer perceptron to input and weight perturbations. IEEE Transactions on Neural Networks. 2001. V. 12. № 6. P. 1358-1366. DOI: 10.1109/72.963772
9. Zimina S.V. Statisticheskiye kharakteristiki iskusstvennoy neyronnoy seti s diskretnym gradiyentnym algoritmom nastroyki s uchetom fluktuatsiy vesovykh koeffitsiyentov. Neyrokompyutery: razrabotka i primeneniye. 2006. № 10. S. 9–15 (in Russian).
10. Zimina S.V. Vliyaniye fluktuatsiy vesovykh koeffitsiyentov na statisticheskiye kharakteristiki iskusstvennoy neyronnoy seti s algoritmom rekurrentnogo obrashcheniya vyborochnoy otsenki korrelyatsionnoy matritsy vkhodnykh signalov. Neyrokompyutery: razrabotka i primeneniye. 2007. № 5. S. 3–7 (in Russian).
11. Litvinov O.S., Zimina S.V. Statisticheskiy analiz fluktuatsiy vesovykh koeffitsiyentov iskusstvennoy neyronnoy seti. nastraivayushcheysya po algoritmu Khebba. Neyrokompyutery: razrabotka i primeneniye. 2009. № 3. S. 33–43 (in Russian).
12. Zimina S.V. Fluktuatsii vesovykh koeffitsiyentov v iskusstvennoy neyronnoy seti s algoritmom Khebba. Neyrokompyutery: razrabotka i primeneniye. 2013. № 4. S. 3– 8 (in Russian).
13. Zimina S.V. Opredeleniye poter v vykhodnom otnoshenii signal/shum adaptivnykh antennykh reshetok. vyzvannykh nalichiyem fluktuatsiy vesovogo vektora. Radiolokatsiya. Rezultaty teoreticheskikh i eksperimentalnykh issledovaniy. Monografiya. V 2-kh knigakh. Kn.1. Pod red. V.D. Yastrebova. M.: Radiotekhnika. 2018. C. 112–118 (in Russian).
14. Zimina S.V. Analiz statisticheskikh kharakteristik adaptivnoy antennoy reshetki s nelineynoy funktsiyey v tsepi korrelyatsion-noy obratnoy svyazi. Radiotekhnika i elektronika. 2005. T. 50. № 8. S. 952–960 (in Russian).
15. Zimina S.V. Fluktuatsii vesovogo vektora v adaptivnykh antennykh reshetkakh s nelineynoy funktsiyey v tsepi obratnoy svyazi. nastraivayushchikhsya po algoritmu rekurrentnogo obrashcheniya vyborochnoy otsenki korrelyatsionnoy matritsy vkhodnykh signalov. Izv. VUZov. Radiofizika. 2006. T. 49. № 2. S. 164–173 (in Russian).
16. Litvinov O.S., Zimina S.V. Analiz vliyaniya fluktuatsiy vesovykh koeffitsiyentov na statisticheskiye kharakteristiki adaptiv-noy antennoy reshetki. nastraivayushcheysya po algoritmu Khebba. Radiotekhnika i elektronika. 2009. T. 54. № 4. S. 423–432 (in Russian).
17. Kruglov V.V., Borisov V.V. Iskusstvennyye neyronnyye seti. Teoriya i praktika. Izd. 2-e M.: Goryachaya liniya – Telekom. 2002. 382 s (in Russian).