Real Spike: Learning Real-Valued Spikes for Spiking Neural Networks

Brain-inspired spiking neural networks (SNNs) have recently drawn more and attention due to their event-driven energy-efficient characteristics. The integration of storage computation paradigm on neuromorphic hardwares makes SNNs much different from Deep Neural Networks (DNNs). In this paper, we argue that may not benefit the weight-sharing mechanism, which can effectively reduce parameters improve inference efficiency in DNNs, some hardwares, assume an SNN with unshared convolution kernels could perform better. Motivated by assumption, a training-inference decoupling method for named as Real Spike is proposed, only enjoys both binary spikes inference-time but also maintains shared Real-valued Spikes during training. This mechanism realized re-parameterization technique. Furthermore, based training-inference-decoupled idea, series forms implementing levels are presented, enjoy convolutions friendly non-neuromorphic hardware platforms. A theoretical proof given clarify Spike-based network superior its vanilla counterpart. Experimental results show all versions consistently performance. Moreover, proposed outperforms state-of-the-art models non-spiking static datasets.