Training Low-Latency Spiking Neural Network with Orthogonal Spiking Neurons

Spiking Neural Networks (SNNs) as a biology-inspired method mimicking the signal transmission of brain neurons can be managed in an event-driven manner, and generate binary signals that are computing-friendly to hardware by their recurrent time window. However, training high-performing SNNs from scratch emerges intrinsic defects, a large number of time steps (>100) is unable to balance the performance with latency. To overcome this issue, we show that SNNs can be trained in fewer time steps (<4) through the proposed Alterable Space-Time LIF (AST-LIF) block, which transforms partial time steps into the space domain. AST-LIF iterates the signals in horizontal and vertical directions in a time step unit, and utilizes the global dependency as Long short-term memory (LSTM) model in a space step unit. Our approach aims to significantly reduce SNNs latency while ensuring high performance. For proof-of-concept, extensive experiments show that balanced spatio-temporal architecture can maintain the performance of SNNs with smaller time steps $(t=2)$.