LSTM neural network implementation using memristive crossbar circuits and its various topologies

dc.contributor.authorAdam, Kazybek
dc.date.accessioned2019-03-11T11:48:52Z
dc.date.available2019-03-11T11:48:52Z
dc.date.issued2018-12
dc.description.abstractNeural Network (NN) algorithms have existed for long time now. However, they started to reemerge only after computers had been invented, because computational resources are required to implement NN algorithms. In fact, computers themselves are not fast enough to train and run the NNs. It can take days to train some complex neural networks for certain applications. One of the complex NNs that became widely used is Long-Short Term Memory (LSTM) NN algorithm. As a broader approach to increase the computation speed and decrease power consumption of neural network algorithms, hardware realizations of the neural networks have emerged. Mainly FPGA and analog hardware are used for these purposes. On this occasion, it happens to be only FPGA implementations of LSTM exist. Using this lack, this thesis work mainly aims to show that LSTM neural network is realizable and functional in analog hardware. In fact, analog hardware using memristive crossbars can be a potential solution to the speed bottleneck experienced in software implementations of LSTM and other complex neural networks in general. This work mainly focuses on implementation of already trained LSTM neural networks in analog circuitry. Since training consists of both forward and backward pass computations through NNs, first, there should be focus on implementing the circuitry that can run forward passes. This forward running circuit further can be extended to a complete circuit which would include training circuitry. Additionally, there exists various LSTM topologies. Software analysis has been done to compare the performance of each LSTM architecture for time-series prediction and time-series classification applications. Each of the architectures can be implemented in analog circuitry without great difficulty using voltage-based LSTM circuit parts due its easiness to reconfigure. Fully functional implementation of the voltage-based memristive LSTM in SPICE circuit simulator is the main contribution of this thesis work. In comparison, current-based LSTM circuit parts may not be easily rearranged due to the difficulty of passing currents from one stage to the next without degradation in magnitude.en_US
dc.identifier.citationAdam, Kazybek (2018). LSTM neural network implementation using memristive crossbar circuits and its various topologies. Nazarbayev University School of Engineeringen_US
dc.identifier.urihttp://nur.nu.edu.kz/handle/123456789/3789
dc.language.isoenen_US
dc.publisherNazarbayev University School of Engineering and Digital Sciencesen_US
dc.rightsAttribution-NonCommercial-ShareAlike 3.0 United States*
dc.rights.urihttp://creativecommons.org/licenses/by-nc-sa/3.0/us/*
dc.titleLSTM neural network implementation using memristive crossbar circuits and its various topologiesen_US
dc.typeMaster's thesisen_US
workflow.import.sourcescience

Files

Original bundle
Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
ThesisFinal_KazybekAdam_Library.pdf
Size:
2.95 MB
Format:
Adobe Portable Document Format
Description:
Thesis
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
6 KB
Format:
Item-specific license agreed upon to submission
Description: