DualNet: Efficient Integration of Artificial Neural Network and Spiking Neural Network with Equivalent Conversion

Abstract

In this paper, we propose Dual Neural Network (DualNet), which integrates ANN and SNN into a single network with layer-by-layer combinations, resulting in high energy efficiency and accuracy. We introduce ANN-to-SNN equivalent conversion (ASEC) that converts ANN layers to the arithmetically equivalent SNN layers without any conversion error. This method enables the transfer of activation between SNN patches and ANN patches while maintaining the equivalence. Also, we propose a computational domain selection method to allocate the domain of each input patch between ANN and SNN for computational cost minimization. As a result, DualNets show no accuracy loss compared to ANNs on large-scale datasets such as CIFAR10 (94.13% top-1), CIFAR100 (72.78% top-1), and ImageNet (72.03% top-1), while their computational energy is up to 47.8% lower than ANNs and up to 35.1% lower than SNNs. The experimental results clearly show that DualNets outperform ANNs and SNNs in terms of accuracy and computational cost.