Memristors have been extensively studied as a promising candidate for next generation non-volatile memory technology. More recently, memristors have also become extremely popular in neuromorphic applications because of their striking resemblance to biological synapses. The memristor was firstly proposed conceptually as the fourth basic electric circuit element whose resistance is dependent on the history of electrical stimulation. Physical implementations of memristors are normally based a solid state, nanoscale metal-insulator-metal (MIM) sandwich structure, and the resistance change is achieved by controlling the ion (either cation or anion) redistribution inside the insulating/switching layer. Specifically, a conductive filament can be formed with a high-concentration of metal cations or oxygen vacancies, leading to an increase in device conductance during set, and a decrease in device conductance during reset associated with the annihilation of the filament. Devices based on such resistive switching mechanisms are often termed resistive random-access memory (RRAM) devices, and offer advantages of simple structure, high density, low power, good endurance, etc. for memory and computing applications.In this dissertation, two kinds of memristor devices will be discussed, using Ag2S and WOx as the switching material, respectively. The WOx device allows incremental modulation of the device conductance, and enables efficient hardware emulation of important synaptic learning functions including paired pulse facilitation, sliding threshold effect, rate dependent plasticity and spike timing dependent plasticity (Chapter 3), showing the resemblance between memristors and biological synapses. Neural networks based on the memristor crossbar array have been used to successfully perform image reconstruction tasks based on the sparse coding algorithm (Chapter 2). A 32×32 WOx memristor crossbar array was used for vector-matrix multiplication acceleration, and the device non-ideality effects in the memristor crossbar array on the image reconstruction performance were examined. Additionally, interesting short-term decay dynamics can be observed in both Ag2S and WOx based devices. Different from the requirements of non-volatile memory which aims for long term memory storage, the volatile nature of these devices can be used to directly encode and process temporal information. Specifically, the Ag2S memristor can encode different neural spiking information in the temporal domain into analog switching probability distributions (Chapter 5). These devices are termed ;;dynamic memristors” and can be applied in novel computing schemes such as reservoir computing systems for efficient temporal information processing including speech recognition (Chapter 4). Both devices show very promising properties for neuromorphic computing – overcoming the von-Neumann bottleneck by incorporating information processing into memory storage. It is believed in the future, very efficient neuromorphic computing chips can be designed and implemented using these memristors that offer potential advantages in terms of area consumption, computing speed and power consumption.
PDF
download
878987797
028-85220240
