Design of Neuromemristive Systems for Visual Information Processing

Neuromemristive systems (NMSs) are brain-inspired, adaptive computer architectures based on emerging resistive memory technology (memristors). NMSs adopt a mixed-signal design approach with closely-coupled memory and processing, resulting in high area and energy efficiencies. Previous work suggests that NMSs could even supplant conventional architectures in niche application domains such as visual information processing. However, given the infancy of the field, there are still several obstacles impeding the transition of these systems from theory to practice. This dissertation advances the state of NMS research by addressing open design problems spanning circuit, architecture, and system levels. Novel synapse, neuron, and plasticity circuits are designed to reduce NMSs’ area and power consumption by using current-mode design techniques and exploiting device variability. Circuits are designed in a 45 nm CMOS process with memristor models based on multilevel (W/Ag-chalcogenide/W) and bistable (Ag/GeS2/W) device data. Higher-level behavioral, power, area, and variability models are ported into MATLAB to accelerate the overall simulation time. The circuits designed in this work are integrated into neural network architectures for visual information processing tasks, including feature detection, clustering, and classification. Networks in the NMSs are trained with novel stochastic learning algorithms that achieve 3.5 reduction in circuit area, reduced design complexity, and exhibit similar convergence properties compared to the least-mean-squares algorithm. This work also examines the effects of device-level variations on NMS performance, which has received limited attention in previous work. The impact of device variations is reduced with a partial on-chip training methodology that enables NMSs to be configured with relatively sophisticated algorithms (e.g. resilient backpropagation), while maximizing their area-accuracy tradeoff