Design of Artificial Neural Networks (ANN) with Domain-Wall Synapse Devices

Having discussed the theory of Artificial Neural Networks (ANNs) and their crossbar-array-based implementations and also the design of domain-wall synapse devices in the previous chapters, here we combine these ideas and discuss implementations of crossbar-array-based ANNs using domain-wall synapse devices. First, we study the system through behaviour-level modelling and analyse the impact of device non-idealities on the performance of the system. We discuss the thresholding-based modification of the standard ANN-training (back-propagation) algorithm for this purpose. Next, we show the design of a single synapse cell in the crossbar array and then the overall arrangement of crossbar arrays such that on-chip learning ability can be incorporated into the system. We support our design with circuit-level SPICE simulations, where the domain-wall synapse device is included as a Verilog A model.