Memristor-based Architectures for Neuromorphic Computing

点击翻译按钮获取中文摘要

The objectives of the proposed research are: (a) To further develop the state-of-the-art for Memristive-based architectures and circuits for various applications such as neuromorphic computing, finite field multipliers, digital filters, computer arithmetic etc. This requires development of various tools such as cell library; accurate modeling of Memristors fabricated using different technologies and better architectures for Memristors Crossbar Arrays. Memristive neural networks have exhibited problems, such as, device non-uniformity, resistance level instability, sneak path currents, and wire resistance. In addition, potential learning algorithms must be able to deal with statistical variations and fluctuations in programmed conductance states and the lack of linear and symmetric responses to electric pulses. Our aim is to study these problems and find solutions for some. We also aim to develop an area efficient Mirrored Memristive Crossbars architecture by reducing the transistors required in that architecture. Among other application for this device we will investigate Memristor-based implementation of Spiking Neural Networks. This emerging device has been regarded as the viable technology for Nano-electronic circuits. Research in this area is very important for Canada to stay competitive in world and keeps its place as leader. (b) Spiking Neural Networks (SNNs) have emerged recently as a good potential for wide ranges of applications like pattern recognition, clustering etc. as it is demonstrated that these networks resemble closely with the activities and function of the brain. There are many models reported in the literature to precisely define their behavior and the functionality of their neurons. These models are divided into three categories namely; Biologically-plausible models, Biologically-inspired models, and High-level models. These models are mostly very complicated for practical implementation. We are aiming to develop accurate and efficient model for digital implementation of SNNs which closely approximate models presented in the literature. We would look at trade off for resource requirements versus higher accuracy that enable the designer to choose the model that fits his/her application. We are also looking at the novel ways of training these SNNs through Spike Time Dependent Plasticity (STDP) for applications such as pattern recognition. Ultimately, research on the design of Memristor-based SNN along with its learning will be carried out. This research is contingent upon developing competitive low-cost, low-power, high-speed and area efficient digital signal processing algorithms and architectures that enable synergy and has the potential to enhance Canada's competitiveness in the area of smart security systems.