Memristor: Modeling, Simulation and Usage in Neuromorphic ComputationHu, Miao (2014) Memristor: Modeling, Simulation and Usage in Neuromorphic Computation. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractMemristor, the fourth passive circuit element, has attracted increased attention from various areas since the first real device was discovered in 2008. Its distinctive characteristic to record the historic profile of the voltage/current through itself creates great potential in future circuit design. Inspired by its high Scalability, ultra low power consumption and similar functionality to biology synapse, using memristor to build high density, high power efficiency neuromorphic circuits becomes one of most promising and also challenging applications. The challenges can be concluded into three levels: device level, circuit level and application level. At device level, we studied different memristor models and process variations, then we carried out three independent variation models to describe the variation and stochastic behavior of TiO2 memristors. These models can also extend to other memristor models. Meanwhile, these models are also compact enough for large-scale circuit simulation. At circuit level, inspired by the large-scale and unique requirement of memristor-based neuromorphic circuits, we designed a circuit simulator for efficient memristor cross-point array simulations. Out simulator is 4~5 orders of magnitude faster than tradition SPICE simulators. Both linear and nonlinear memristor cross-point arrays are studied for level-based and spike-based neuromorphic circuits, respectively. At application level, we first designed a few compact memristor-based neuromorphic components, including ``Macro cell'' for efficient and high definition weight storage, memristor-based stochastic neuron and memristor-based spatio temporal synapse. We then studied three typical neural network models and their hardware realization on memristor-based neuromorphic circuits: Brain-State-in-a-Box (BSB) model stands for level-based neural network, and STDP/ReSuMe models stand for spiking neural network for temporal learning. Our result demonstrates the high resilience to variation of memristor-based circuits and ultra-low power consumption. In this thesis, we have proposed a complete and detailed analysis for memristor-based neuromorphic circuit design from the device level to the application level. In each level, both theoretical analysis and experimental data versification are applied to ensure the completeness and accuracy of the work. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}