Abstract
Realizing the potential of wireless edge artificial intelligence (AI) requires training complex models on resource-constrained devices using diverse datasets. This paper introduces Lead Federated Neuromorphic Learning (LFNL), a decentralized, energy-efficient framework based on Spiking Neural Networks (SNNs). LFNL allows edge devices to collaboratively train global models using a brain-inspired architecture while preserving privacy. Experimental results demonstrate that LFNL achieves comparable accuracy to standard edge AI while reducing data traffic by >3.5×, computational latency by >2.0×, and energy consumption by >4.5×.
Introduction
Edge AI faces significant hurdles: individual devices often lack sufficient local data (data scarcity), and their limited battery life and processing power make training standard Artificial Neural Networks (ANNs) difficult. While Federated Learning (FL) addresses data scarcity by sharing model updates, it typically relies on a central server and energy-intensive ANNs. LFNL addresses these challenges by combining neuromorphic computing with a decentralized leader-follower architecture.
Methodology
LFNL integrates two core innovations to optimize edge computing:
Neuromorphic Computing with SNNs
Unlike traditional ANNs that use continuous values, SNNs process information via discrete temporal "spikes." This sparse computing model mimics biological neurons, replacing power-hungry multiplications with simple additions. This allows for "sparse" computing—if there is no spike, there is no computation—leading to significant energy savings. To overcome the non-differentiable nature of spikes, a pseudo-gradient method is used for backpropagation training.
Lead Federated Learning
To eliminate the need for a fixed central server, LFNL employs a leader election scheme. A device with superior resources (e.g., higher battery or signal strength) is elected as the "leader" to aggregate model weights from "follower" peers. This decentralized approach avoids single points of failure and reduces communication bottlenecks often found in fully peer-to-peer mesh networks.
Experimental Results
The framework was evaluated across several modalities, including audio (traffic sounds), visual (traffic images), and radar (gesture recognition), as well as standard benchmarks like CIFAR-10/100.
- Energy and Communication Efficiency: LFNL reduced energy consumption by over 4.5× and data traffic by over 3.5× compared to standard FL.
- Computational Speed: Processing latency was halved, achieving a >2.0× reduction.
- Accuracy and Robustness: Accuracy remained high (e.g., 95.6% for visual tasks), with only a marginal 1.5% loss compared to centralized ANN training. Crucially, the system proved stable in "Non-IID" scenarios, where data distribution across devices is highly uneven.
Conclusion
LFNL provides a scalable, ultra-low-power solution for decentralized edge AI. By leveraging the sparsity of spiking neural networks and a flexible leader-follower architecture, it overcomes the primary energy and communication barriers of wireless edge computing. Future work will focus on narrowing the accuracy gap and improving leader resilience in dynamic environments where devices may frequently move out of range.
Summary updated