A Brain-inspired Fully Hardware Hopfield Neural Network based on Memristive Arrays

In this work, we propose a fully hardware implementation of Hopfield neural network (HNN) based on memristive arrays, which adopts a bottom-up brain-inspired hardware framework. Memristors are used as key computing units to design cell modules in the low layer to play roles similar to neurons and synapses, so as to perform information integration, filtering, and plasticity of weights in a simple circuit structure and in-memory computing way. Cell modules are cascaded by the mode of encoding and mapping based on HNN in a structured regular circuit way to construct functional modules with brain-inspired parallel analog computing capacity in middle layers. Different functional modules perform information interaction according to the system's requirements and goals, thus realizing the overall system in the top layer. Our proposed HNN system is then used to solve combinatorial optimization problems. Different from other similar work, our system starts from a memristor-based brain-inspired framework and is implemented fully in hardware. The experimental results show that our work can not only improve the convergence speed, but also can be conveniently used to solve problems of different scales because of its good scalability. In addition, with hardware overhead and power consumption analysis, our system has been shown to be very hardware-friendly. Our work represents an advance towards a memristor-based hardware system with brain-inspired structure and high performance.