A fast and ultra low power time-based spiking neuromorphic architecture for embedded applications

Time-based Spiking Neural Network (SNN) has recently received increased attentions in neuromorphic computing system designs due to more bio-plausibility and better energy-efficiency. However, unleashing its potentials in realistic cognitive applications is facing significant challenges such as inefficient information representations and impractical learnings. In this work, we aim for exploring a practical time-based Spiking Neuromorphic Engine (SNE) to fulfill the demand of real-world applications. A holistic hardware-favorable solution set across time-based coding, learning and decoding is proposed accordingly to close the gap between hardware and bio-plausibility. Experimental results in cognitive benchmarks (e.g. MNIST dataset) show that our proposed SNE achieves remarkable improvements in synaptic efficiency and power with a comparable accuracy and throughput when compared to the popular rate-coding based SNN and Artificial Neural Network (ANN). Unlike the complicated convolutional neural network (CNN) or deep neural network (DNN) that requires expensive hardware resource, our work prototypes a light but powerful time-based SNN framework with unique advantages for cognitive tasks performed in ultra low power and resource constrained platforms.