Artificial Neural Networks Modeling Using Memristors for Small Physical Devices

Small physical device like IoT devices require sensing nodes equipped with wireless communication that will communicate vital information from sensing nodes to the cloud. Given that energy is very limited at the sensing node, most of the IoT nodes include very limited processing and hence this vital information is not processed and transmitted as crude data and as a result causes an increase in the power and bandwidth, two features that are scarce in IoT infrastructure, particularly at the sensing end. This project aims to redesign the conventional architecture by reallocating parts of the intelligence from the centralized system to these small physical devices to allow them to perceive, plan and execute autonomously. However, such designs require the implementation of on chip Artificial Intelligence (AI) training and inference mechanisms that uses very small physical space and consumes very tiny power. This requires implementing AI algorithms and deep learning using compact architectures and ultra-low power implementation techniques. With the discovery of the memristors, researchers were able to model and implement Artificial Neural Networks (ANN) on small integrated circuits consuming reasonably low power. But memristors are still challenging and not yet mature given the many drawbacks. In this research project, we plan to address on the fundamental problems of implementing ANN using the memristor crossbar architecture which is sneak path current. In this research we proposed and evaluated a solution that is able to overcome the problem of sneak path by using each memristor crossbar as a single weight in the network. We have also identified a gap in literature in terms of considering the sneak path current in crossbar architectures in the context of ANN which was proved through a case study by re-implementing existing work.