Сибирский федеральный университет

Тип публикации: доклад, тезисы доклада, статья из сборника материалов конференций

Конференция: Computer Science On-line Conference, CSOC 2020

Год издания: 2020

Идентификатор DOI: 10.1007/978-3-030-51971-1_35

Ключевые слова: echo state neural networks, field programmable integrated circuits, hyperdimensional calculations, recurrent neural networks

Аннотация: Real-time data processing using recurrent neural networks (NN) is non-trivial task, due to tight timing constraints requirements. It is proposed hardware implementation of recurrent echo state NN (ESN) on the basis of the Cyclone IV FPGA. Advantages of the hardware implementation are high computational parallelism and low power consumption. To solve the problem of neuron weight storage, it is proposed to reduce the space of their values to a set of integers of low capacity. It was determined that the proposed NN model decreases need in hardware resources for the reservoir implementation in 2–3 orders of magnitude in comparison with conventional NN. Modeling results, implementation and testing of the FPGA project confirmed effectiveness of the proposed integer NN in hardware applications #CSOC1120. © 2020, Springer Nature Switzerland AG.

Журнал: Advances in Intelligent Systems and Computing

Выпуск журнала: Vol. 1225 AISC

Номера страниц: 429-437

ISSN журнала: 21945357

Издатель: Springer

• Nepomnyashchiy O. (Siberian Federal University, Svobodny Prosp. 79, Krasnoyarsk, 660041, Russian Federation)
• Khantimirov A. (Siberian Federal University, Svobodny Prosp. 79, Krasnoyarsk, 660041, Russian Federation)
• Galayko D. (Sorbonne Université, Jussieu Campus, Paris, 75006, France)
• Sirotinina N. (Siberian Federal University, Svobodny Prosp. 79, Krasnoyarsk, 660041, Russian Federation)

• Scopus
• Ядро РИНЦ (eLIBRARY.RU)