https://doi.org/10.1140/epjs/s11734-024-01369-y
Regular Article
Automated design of deep neural networks with in-situ training architecture based on analog functional blocks
Saint Petersburg Electrotechnical University “LETI”, Professora Popova St. 5, 197022, Saint Petersburg, Russia
Received:
14
September
2024
Accepted:
10
October
2024
Published online:
24
October
2024
Deep neural networks are the most in demand and popular for solving a wide range of problems in various fields of science and technology due to their learning ability. Increasing the number of layers in software implementation of deep neural networks (DNN) requires increased power consumption when working with big asynchronous data and becomes ineffective. To solve this problem, this study proposes a new approach to DNN design with on-chip training, based on the analog implementation of all computational operations, including matrix–vector multiplication in synaptic crossbar arrays, with digital implementation of synaptic weight setup and storage operations. The efficiency of the proposed approach is determined by the developed integrated synaptic CMOS IP block, which performs direct (without DAC) linear conversion of the binary code of the synaptic weight into the level of synaptic conductivity, as well as unified analog CMOS IP blocks with the ability to change the number of inputs and outputs. The paper presents the circuit diagrams, topology and results of SPICE modeling of CMOS IP blocks, as well as a description of the functional capabilities of the compiler, which allows using these IP blocks to synthesize functional circuits and topology of deep neural networks with specified parameters based on a neural network graph with training using the backpropagation method.
M. O. Petrov, E. A. Ryndin and N. V. Andreeva have contributed equally to this work.
Copyright comment Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
© The Author(s), under exclusive licence to EDP Sciences, Springer-Verlag GmbH Germany, part of Springer Nature 2024. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.