AS-FedBridge bridges ANN-SNN gap for efficient federated learning

Federated learning lets distributed edge devices train models collaboratively without sharing raw data, but deploying it on resource-constrained hardware is tough. Spiking Neural Networks (SNNs) offer high energy efficiency through sparse, event-driven computation, yet they represent information as discrete spatio-temporal spikes—fundamentally different from the continuous real-valued activations of traditional Artificial Neural Networks (ANNs). This semantic gap creates representational misalignment when jointly training both network types, limiting practical mixed-architecture federated systems.

AS-FedBridge solves this with a lightweight Bridge equipped with a Pseudo-Spike Interface that continuously projects ANN signals into a spike-compatible space, enabling effective alignment between heterogeneous clients. The framework consistently outperforms advanced heterogeneous FL methods across four benchmarks, showing a positive correlation between alignment quality and collaborative performance. Crucially, it handles extreme scale, architecture, and client heterogeneity while introducing only marginal computational overhead—giving operators a highly controllable trade-off between model accuracy and resource efficiency on edge devices.