Evaluation of storage and functioning characteristics of artificial neural networks on the basis of a neurocomputer device

A solution is proposed for assessing the temporal and quantitative characteristics of the storage and processing of artificial neural networks based on a neurocomputer device. The most popular and used topologies of artificial neural networks are considered (single-layer and multi-layer perceptron, Hopfield networks, Hamming networks, BAM networks, Jordan networks, Elman networks, ART networks of various modifications, Grossberg star, Kohonen networks, radial basis neural networks, backward propagation networks, convolutional networks) for which analytical relationships are given to evaluate the training cycle of an artificial neural network, the amount of necessary memory and the amount of data transmitted. The difference between the proposed results is that to assess the functioning of artificial neural networks, approaches and characteristics inherent in the class of neuroprocessor devices are offered and only when implementing the calculations presented in the neural network logical basis, which allows to increase the efficiency of the task solution based on neurocomputer devices. An artificial neural network was considered using a set-theoretic approach, which allowed us to obtain analytical relationships based on the number of neuron emulation operations and connections between neurons in accordance with the topology of the neural network.

1. Galushkin A.I. Neirokomp'yutery. Kn. 3: ucheb. Posobie dlya vuzov. IPRZhR. 2000.

2. ZAO NTTs «Modul'». Saitkompanii NTTs «Modul'». 2020. Dostupno po adresu: https://www.module.ru/directions/iskusstvennyj-intellekt (data obrashcheniya: 01.04.2020).

3. Romanchuk V.A. Razrabotka algoritmicheskogo i programmnogo obespecheniya kontrolya funktsionirovaniya slozhnykh, raspredelennykh i oblachnykh system na baze neiroprotsessorov. Modelirovanie, optimizatsiya i informatsionnye tekhnologii. 2015;3(10). Dostupno po adresu: http://moit.vivt.ru/wpcontent/uploads/2015/10/Romanchuk_3_15_1.pdf.

4. Romanchuk V.A., Ruchkin V.N., Fulin V.A. Proektirovanie neiroprotsessornykh system na osnove nechetkoi klasterizatsii. Vestnik RGRTU. 2014;4(50-1):87-93.

5. Romanchuk V.A. Matematicheskoe obespechenie otsenki proizvoditel'nosti system neirokomp'yuternoi obrabotki informatsii. Izvestiya Tul'skogo gosudarstvennogo universiteta. Tekhnicheskie nauki. 2016;2:242-251.

6. Romanchuk V.A. Metod organizatsii parallelizma dannykh v neiroprotsessorakh. Izvestiya Tul'skogo gosudarstvennogo universiteta. Tekhnicheskie nauki. 2018;9:148-157.

7. Romanchuk V.A. Razrabotka algoritma optimizatsii zadachi klasterizatsii raspredelennykh vychislitel'nykh system na baze neiroprotsessorov. Informatika i prikladnaya matematika: mezhvuzovskii sbornik nauchnykh trudov. 2016;22:79-83.