A SPIKING SEQUENTIAL MODEL: RECURRENT LEAKY INTEGRATE-AND-FIRE

Daiheng Gao, Hongwei Wang, Hehui Zhang, Meng Wang, Zhenzhi Wu

Keywords: spiking neural network, RNN, spiking mode, brain-inspired, text summarization, DVS

Abstract: Stemming from neuroscience, Spiking neural networks (SNNs), a brain-inspired neural network that is a versatile solution to fault-tolerant and energy efficient information processing pertains to the ”event-driven” characteristic as the analogy of the behavior of biological neurons. However, they are inferior to artificial neural networks (ANNs) in real complicated tasks and only had it been achieved good results in rather simple applications. When ANNs usually being questioned about it expensive processing costs and lack of essential biological plausibility, the temporal characteristic of RNN-based architecture makes it suitable to incorporate SNN inside as imitating the transition of membrane potential through time, and a brain-inspired Recurrent Leaky Integrate-and-Fire (RLIF) model has been put forward to overcome a series of challenges, such as discrete binary output and dynamical trait. The experiment results show that our recurrent architecture has an ultra anti-interference ability and strictly follows the guideline of SNN that spike output through it is discrete. Furthermore, this architecture achieves a good result on neuromorphic datasets and can be extended to tasks like text summarization and video understanding.