[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Edge Intelligence with Spiking Neural Networks

Created by
  • Haebom

Author

Shuiguang Deng, Di Yu, Changze Lv, Xin Du, Linshan Jiang, Xiaofan Zhao, Wentao Tong, Xiaoqing Zheng, Weijia Fang, Peng Zhao, Gang Pan, Schahram Dustdar, Albert Y. Zomaya

Outline

This paper is written against the backdrop of the growing interest in enabling intelligent services directly on resource-constrained devices due to the convergence of artificial intelligence and edge computing. Existing deep learning models require significant computing resources and centralized data management, and the resulting latency, bandwidth consumption, and privacy concerns expose the critical limitations of the cloud-centric paradigm. In this paper, we provide a comprehensive overview of Edge Intelligence (EdgeSNNs) based on spiking neural networks (SNNs) that mimic the dynamics of biological neurons to achieve low-power event-based computing. We present a systematic taxonomy of EdgeSNNs, including neuron models, learning algorithms, and supporting hardware platforms, and deeply discuss three representative practical considerations: on-device inference using lightweight SNN models, resource-aware learning and updating under abnormal data conditions, and security and privacy concerns. We also highlight the limitations of evaluating EdgeSNNs on existing hardware and introduce a dual-tracking benchmarking strategy to support fair comparison and hardware-aware optimization. This study aims to bridge the gap between brain-inspired learning and real-world edge deployments, and provide insights into current progress, open challenges, and future research directions. This paper is the first dedicated and comprehensive survey of EdgeSNN, and serves as an essential reference for researchers and practitioners working at the intersection of neuromorphic computing and edge intelligence.

Takeaways, Limitations

Takeaways:
A comprehensive investigation of SNN-based edge intelligence (EdgeSNNs) suggests the possibility of building low-power, low-latency, and privacy-enhanced edge computing environments.
Provides a systematic classification of EdgeSNNs' neuron models, learning algorithms, and hardware platforms.
On-device inference using lightweight SNN models, resource-aware learning and updating, with practical considerations for security and privacy issues.
Proposing a dual-tracking benchmarking strategy for fair comparison and hardware-aware optimization.
Provides important reference material for research and development in neuromorphic computing and edge intelligence.
Limitations:
As this is still an early stage research, further research is needed on the performance and stability of EdgeSNNs in real edge environments.
Further research is needed on the generalizability of EdgeSNNs to various edge devices and applications.
New benchmarking methods and hardware development are needed to overcome the limitations of existing hardware evaluations.
Lack of extensive experimental validation across diverse datasets and tasks.
👍