A versatile neuromorphic system based on simple neuron model

Abstract

Brain-inspired neuromorphic computing has attracted much attention for its advanced computing concept. However, the massive hardware cost in fully-connected architectures makes it challenging to build a large-scale neuromorphic system. In this work, we report a compact, programmable, versatile, and scalable neuromorphic architecture. To demonstrate the concept of the neuromorphic architecture, a neuromorphic system consisting of four cores is implemented on an FPGA platform. On the one hand, the neuromorphic system is extremely compact and hardware-saving. The computing block based on a simple digital leaky Integrate-and-Fire (LIF) model only costs 69 logic elements (LEs); only one physical neuron is implemented in each core, and it can be reused as hundreds of virtual neurons by time-division-multiplexing; only four 9-bit synaptic weights are assigned to each neuron, which effectively alleviates the hardware explosion in fully-connected architecture. On the other hand, the neuromorphic system is programmable and versatile, and can perform different neural network computing. The neuromorphic system mapped with a three-layer feedforward network successfully recognizes the MNIST handwritten digits with an accuracy of 96.26%, and it also effectively realizes different convolution operations which are basic computing operations in convolutional neural networks. Last but not least, each neuromorphic core has its own router module, making it convenient to scale up.