Memristor-Based Neuromorphic Computing and Artificial Neural Networks for Computer Vison and AI—Applications

The traditional von Neumann architecture seen in digital processors generally suffers from the data transfer rate bottleneck and inefficiency in terms of energy consumption. The objective is to emulate the computational functions of the human brain, with special attention to attaining high compactness and energy efficiency. On the way to achieving this goal, there are a number of formidable obstacles. Because the biological brain uses extremely low power (approx. 40–60 W), high-density neural networks. Therefore new device concepts that combine low power consumption, scalability, and advanced computing functionality are needed. In this article, we will try to summarize the memristive-based neuromorphic computing. Here, the physics and functioning of CMOS-based floating-gate memory devices in artificial neural networks are explored, and then different memristive concepts relevant to deep neural network and spiking neural network architectures are reviewed and discussed. The final section discusses the main technological obstacles and potential directions for neuromorphic computing.