Designing Neuromorphic Computing Systems with Memristor DevicesMahmoud, Amr (2019) Designing Neuromorphic Computing Systems with Memristor Devices. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractDeep Neural Networks (DNNs) have demonstrated fascinating performance in many real-world applications and achieved near human-level accuracy in computer vision, natural video prediction, and many different applications. However, DNNs consume a lot of processing power, especially if realized on General Purpose GPUs or CPUs, which make them unsuitable for low-power applications. On the other hand, neuromorphic computing systems are heavily investigated as a potential substitute for traditional von Neumann systems in high-speed low-power applications. One way to implement neuromorphic systems is to use memristor crossbar arrays because of their small size, low power consumption, synaptic like behavior, and scalability. However, these systems are in their early developing stages and still have many challenges to be solved before commercialization. In this dissertation, we will investigate designing of neuromorphic computing systems, targeting classification and generation applications. Specifically, we introduce three novel neuromorphic computing systems. The first system implements a multi-layer feed-forward neural network, where memristor crossbar arrays are utilized in realizing a novel hybrid spiking-based multi-layered self-learning system. This system is capable of on-chip training, whereas for most previously published systems training is done off-chip. The system performance is evaluated using three different datasets showing improved average failure error by 42% than previously published systems and great immunity against process variations. The second system implements an Echo State Network (ESN), as a special type of recurrent neural networks, by utilizing a novel memristor double crossbar architecture. The system has been trained for sample generation, using the Mackey-Glass dataset, and simulations show accurate sample generation within a 75% window size of the training dataset. Finally, we introduce a novel neuromorphic computing for real-time cardiac arrhythmia classification. Raw ECG data is directly fed to the system, without any feature extraction, and hence reducing classification time and power consumption. The proposed system achieves an overall accuracy of 96.17% and requires only 34 ms to test one ECG beat, which outperforms most of its counterparts. For future work, we introduce a preliminary neuromorphic system implementing a deep Generative Adversarial Network (GAN), based on ESNs. The system is called ESN-GAN and it targets natural video generation applications. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}