Automatic generation of spiking neural networks on neuromorphic computing hardware for IoT edge computing

초록

Recently, there has been a remarkable increase in the demand for edge computing on the Internet of Things (IoT), which encompasses a network of interconnected devices capable of operating independently and making intelligent decisions. To enable low-power and real-time processing, spiking neural networks (SNNs) on neuromorphic computing hardware are expected to emerge as a prominent solution for supporting IoT edge computing. SNN models need to be optimized to meet the specific quality of service requirements according to a variety of IoT edge services. However, the complex dynamics and non-differentiable nature of SNNs are challenging issues for implementing the SNN models. The heterogeneity of neuromorphic computing hardware makes it more difficult to generate SNN models that meet specified requirements, such as accuracy or execution time. Existing IoT platforms support various neural functionalities, yet they are typically unsuitable for generating SNN models customized to meet user performance requirements for IoT edge computing, especially considering the neuromorphic computing hardware. In this paper, we propose an automatic SNN generation technique on neuromorphic computing hardware for IoT Edge computing. The goal is to automatically generate SNN models that efficiently balance cost while satisfying user-specified performance requirements and hardware constraints. Our proposed approach enables rapid prototyping by using a novel predictive model based on profiling results obtained from FPGA-based hardware, as well as neuromorphic hardware like Loihi. We built its implementation that is compatible with the open IoT platform Node-RED and measured its performance. The experimental results show that our approach enables the automatic generation of SNN models with an average requirement meeting ratio of over 96%, demonstrating its efficacy for IoT edge computing. Furthermore, a voice recognition service implementation in Node-RED demonstrated the feasibility of our approach, confirming that SNN models generated through our framework can effectively operate on FPGA-based neuromorphic hardware within IoT platforms.

키워드