Nonpolar Neuron for ANN-SNN Conversion Toward Ternary Spiking Neural Network

Spiking neural network (SNN) has been widely acknowledged as a promising approach to enabling efficient neuromorphic computing. Compared to direct training of SNN, the conversion of artificial neural network (ANN) to SNN can fully leverage the existing techniques in ANN, thereby preserving the superior performance of deep learning algorithms. However, common challenges of high latency, excessive power consumption or reduced accuracy were frequently met by previous conversion methods. To address these issues, we propose a novel nonpolar neuron model for converting ANN to Ternary Spiking Neural Network (TSNN), capable of emitting both positive and negative spikes and thereby owning higher information encoding density. We demonstrate our TSNN on complex visual recognition task with nearly lossless accuracy with respect to ANN counterpart. In addition, we proposed a hardware implementation approach for TSNN using analog-digital hybrid circuits. Our results highlight remarkable computational efficiency of SNN inference.