Loading...
Citations
Altmetric:
Abstract
Memristor or RRAM (Resistive Random-Access Memory) based crossbar array architecture is considered a leading contender for the next-generation non-volatile memory (NVM) as well as for future computing paradigms, such as in-sensor computing, neuromorphic computing, neural networks, analog computing, reconfigurable computing, stochastic computing etc. Traditional Artificial Neural Networks are very useful for solving nonlinear problems but are not designed to solve temporal problems. Recurrent Networks are suitable for solving temporal problems, but they involve complicated circuits and are costly to train. Reservoir Computing is a computing paradigm that is specialized at solving temporal problems with less training cost. Reservoir Computing is a framework that can extract features from temporal inputs and cast them into a higher dimension feature space. The reservoir is followed by a readout layer that can analyze the extracted features to accomplish tasks such as inference for classification. Reservoir Computing systems possess an advantage since the training is only performed at the readout layer and hence can process complicated temporal data with a low training cost. Three of the main requirements for a reservoir include a high nonlinearity, a fading memory of past operations which would lead to the decay of such memory over time. Diffusive memristors are inherently nonlinear and have dynamical memories. Hence, diffusive memristors are suitable as units of a reservoir layer. The readout layer can be implemented using nonvolatile drift memristors in a 1-Transistor 1-Memristor(1T1R) format by harnessing its ability to efficiently compute Multiply Accumulate Operations by harnessing Kirchhoff’s Current Law. In my thesis, we have experimentally implemented a physical Reservoir Computing system using diffusive memristor based reservoir and drift memristor based readout layer. The rich nonlinear dynamic behavior possessed by a diffusive memristor owing to Ag migration in dielectrics and the robustness of in situ training of 1T1R makes the system ideal for temporal pattern classification. We then demonstrated experimentally that our reservoir computing system can successfully identify handwritten digits from the MNIST dataset, achieving an accuracy of 83%. We are also designing and about to fabricate a few other candidate memristor devices that can potentially improve the effectiveness of the reservoir layer and at the same time be faster as well as consume less energy. We also explored the possibility of using the devices developed above for performing stochastic computing. Stochastic computing is a computing paradigm where the probability of 1’s occurring in a bit stream carries the information. One of the main advantages of this kind of computing is its resilience to bit-flipping noise in performing multiplications. While the output signal of a sensor is typically in the form of analog amplitude the biological neural networks process information represented by the timing of a spike. Converting a signal of an analog amplitude into the timing of a spike needs bulky CMOS circuits. We showed that such conversion can be efficiently achieved using a single diffusive memristor. The delay time of diffusive memristors follows a clear trend where it increases with the magnitude of the pulse applied. We also propose on harnessing this property in performing bio-realistic learning such as STDP.
Type
campusfive
dissertation
dissertation
Date
2022-05
Publisher
Degree
Advisors
License
License
http://creativecommons.org/licenses/by-nc-nd/4.0/