S.N. Danilin – Ph. D. (Eng.), Associate Professor,

Murom Institute (branch) Federal state budgetary Educational Institution of Higher Education 

«Vladimir State University named after A.G. and N.G. Stoletovs»

E-mail: dsn-55@mail.ru

S.A. Shchanikov – Ph. D. (Eng.), Associate Professor, Dean, The faculty of Information Technology, Murom Institute (branch) 

Federal state budgetary Educational Institution of Higher Education 

«Vladimir State University named after A.G. and N.G. Stoletovs»

E-mail: seach@inbox.ru

I.A. Bordanov – engineer,

Murom Institute (branch) Federal state budgetary Educational Institution of Higher Education 

«Vladimir State University named after A.G. and N.G. Stoletovs»

E-mail: bordanov2011@yandex.ru

A.D. Zuev – engineer,

Murom Institute (branch) Federal state budgetary Educational Institution of Higher Education  «Vladimir State University named after A.G. and N.G. Stoletovs» E-mail: ad-nemo@mail.ru

The authors devote this article to the theory of computer-aided design of ANNM with a given fault tolerance (FT), in terms of the development, research and application of a system quantitative criterion (measure) for determining FT ANNM on an arbitrary implementation platform. A quantitative criterion will determine the fault tolerance levels of various ANNM options of arbitrary purpose and complexity.

In published works devoted to the problem of ensuring high fault tolerance, ANNs considered, posed and solved many problems in this area.

It is shown that work in the field of FT ANN is carried out for a long time by many researchers. However, the problem is complex, multifaceted, with increasing dimension when scaling computational structures, and therefore is solved slowly and fragmented, requiring the use of additional resources in solving various computational problems.

In the well-known works, the features of ANNM, as single physical-informational objects, are practically not affected. The proposed criteria for assessing FT are not systemic (not based on the theory of system analysis), are qualitative or indirect and are not consistent with current regulatory documents on the reliability of software and hardware systems.

A general approach to solving the problem is proposed and justified. It is based on the theory of systems analysis and includes the following main points and assumptions: ANNM are unified physical-informational, hardware-software trained objects; the above components of ANNM have a joint, generally dependent effect on all their parameters and characteristics; ANNM must be investigated, designed, manufactured and operated as a single physical and information objects implemented by hardwaresoftware learning tools; modern ANNM and technical tools based on nanomemristors of a practical level of complexity, factors that destabilize their work, as well as the tasks they solve – difficult to formalize or not formalized; design and research of ANNM is successfully carried out by methods of simulation of information processes and systems adapted for ANNM.

The requirements for the FT criterion are substantiated: systematic and quantitative; invariance to the type, structure and parameters of ANNM; consistency with standards in the design of technical facilities; consistency with standards in the field of reliability of technical facilities; dependence on operating time.

The analysis of terminology in the current standards in the field of reliability is carried out, significant differences between them are shown. The definition of FT ANNM is formulated and justified.

The relationship of accuracy, FT, and ANNM reliability is shown.

The need for the inclusion of the functional tolerances of ANNM in the structure of the criterion for determining FT is shown. Based on the general approach and the determination of FT, a variant of its criterion U(t,v) is synthesized, which is a function of the operating time – the duration and (or) the amount of work performed by the ANNM.

An example of the technology for applying the developed criterion at the stage of preparation for the technical implementation of ANNM based on arrays of metal-oxide memristive devices in the cross-bar topology of the test complexity level is given.

The practical application of the criterion U(t,v) allows us to quantify the FT of arbitrary ANNMs and effectively improve the known methods and develop new methods for passively and (or) actively providing this property.

Danilin S.N., Shchanikov S.A., Bordanov I.A., Zuev A.D. Quantitative determination of fault tolerance of artificial neural networks based on memristors. Neurocomputers. 2020. V. 22. № 1. P. 55–65. DOI: 10.18127/j19998554-202001-06.

1. Danilin S., Shchanikov S., Bordanov I., Zuev A., Sakulin A. The system approach to designing memristor-based neural networks. Proc. MEMRISYS. 2019. P. 1.
2. Schuman C.D., Potok T.E., Patton R.M., Birdwell J.D., Dean M.E., Rose G.S., Plank J.S. A Survey of Neuromorphic Computing and Neural Networks in Hardware. Proceedings of CoRR journal. [Elektronnyy resurs] URL: https://arxiv.org/pdf/1705.06963.pdf (Data obrashcheniya: 24.08.2019).
3. Chekushkin V.V., Zhiganov S.N. Computational methods in optimization of engineering problems. Raleigh. North Carolina USA: Open Science Publishing. 2018. P. 202.
4. Starovoytov A.V., Galushkin A.I. Novyye tekhnologii mikroelektroniki i razrabotki perspektivnykh neyrokompyuterov. Informatizatsiya i svyaz. 2017. №1. S. 7-17. (in Russian).
5. Torres-Huitzil C., Girau B. Fault and error tolerance in neural networks: A review. IEEE Access. 2017. V. 5.P. 17322-17341.
6. Yeung D.S., Cloete I., Shi D., Ng W.W.Y. Sensitivity Analysis for Neural Networks. Heidelberg: Springer. 2010. P. 89.
7. Merritt R. AI Becomes the New Moore’s Law // EE Times. [Elektronnyy resurs] URL: https://www.eetimes.com/document.asp?doc_id=1333471 (Data obrashcheniya: 3.09.2019).
8. Galushkin A.I., Pantyukhin D.V. SuperEVM i memristory. Informatsionnyye tekhnologii. 2016. T. 22. № 4. S. 304-312  (in Russian).
9. Mikhaylov A.N., Morozov O.A., Ovchinnikov P.E., Antonov I.N., Belov A.I., Korolev D.S., Sharapov A.N., Gryaznov E.G., Gorshkov O.N., Pigareva Y.I., Pimashkin A.S., Lobov S.A., Kazantsev V.B. One-board design and simulation of double-layer perceptron based on metal-oxide memristive nanostructures. IEEE Transactions on Emerging Topics in Computational Intelligence. 2018. V. 2. № 5. P. 371-379.
10. Merrikh F., Prezioso M., Chakrabarti B., Kataeva I., Strukov D. Implementation of multilayer perceptron network with highly uniform passive memristive crossbar circuits. Nature Communications. 2018. V. 9. № 1.
11. Kataeva I., Ohtsuka S., Nili H., Kim H., Isobe Y., Yako K., Strukov D. Towards the development of analog neuromorphic chip prototype with 2.4m integrated memristors. IEEE International Symposium on Circuits and Systems (ISCAS). 2019. P. 1-5.
12. Nikiruy K.E., Emelyanov A.V., Rylkov V.V., Sitnikov A.V., Demin V.A. Adaptive Properties of Spiking Neuromorphic Networks with Synapses Based on Memristive Elements. Technical Physics Letters. 2019. V. 45. P. 386-390.
13. Galushkin A.I. Neyronnyye seti: osnovy teorii. M.: Goryachaya liniya-Telekom. 2013. 496 s. (in Russian).
14. Lanza M., Wong P., Pop E., Ielmini D. Recommended Methods to Study Resistive Switching Devices. Advanced Electronic Materials. 2019. V. 5. P. 1-28.
15. Chua L. Five non-volatile memristor enigmas solved. Applied Physics. 2018. V. 124. P. 563-606.
16. Mehonic A., Joksas D., Ng W., Buckwell M., Kenyon A. Simulation of Inference Accuracy Using Realistic RRAM Devices Front. Neuroscience. 2019. V. 13. R. 1-15.
17. Galushkin A.I., Danilin S.N., Shchanikov S.A. The research of memristor-based neural network components operation accuracy in control and communication systems. Proceedings of International Siberian Conference on Control and Communications. 2015. P. 1-6.
18. GOST 27.002-2015 Nadezhnost v tekhnike. Terminy i opredeleniya. Ofitsialnoye izdaniye. M.: Standartinform. 2016. 28 s.
19. Details for IEV number 192-01-22: «dependability». IEC 60050 International Electrotechnical Vocabulary. [Elektronnyy resurs] URL: http://www.electropedia.org/iev/iev.nsf/display?openform&ievref=192-01-22 (data obrashcheniya: 23.01.2019).
20. Details for IEV number 192-10-09: «fault tolerance». IEC 60050 – International Electrotechnical Vocabulary. [Elektronnyy resurs] URL: http://www.electropedia.org/iev/iev.nsf/display?openform&ievref=192-10-09 (data obrashcheniya: 23.01.2019).
21. GOST R 27.301–2011. Nadezhnost v tekhnike. Upravleniye nadezhnostyu. Tekhnika analiza bezotkaznosti. Osnovnyye polozheniya. M.: Standartinform. 2013. 19 s. (in Russian).
22. Danilin S.N., Shchanikov S.A., Panteleev S.V. Determining Operation Tolerances of Memristor-Based Artificial Neural Networks. Engineering and Telecommunication (EnT). 2016. P. 34-38.
23. Danilin S.N., Shchanikov S.A., Sakulin A.E., Bordanov I.A. Determining the Fault Tolerance of Memristors-Based Neural Network Using Simulation and Design of Experiments. Engineering and Telecommunication (EnT-MIPT). 2018. S. 205-209.
24. Morozov O.A., Ovchinnikov P.E., Mikhaylov A.N. Razrabotka iskusstvennoy neyronnoy seti s formirovaniyem vesovykh koeffitsiyentov na osnove memristivnykh nanostruktur. Trudy XXI nauch. konf. po radiofizike. 2017. S. 348-350. (in Russian).