Title: SpikeWFM: A Spiking-Aided Wireless Foundation Model for Resilient Channel Prediction

Abstract:

This study introduces SpikeWFM, an innovative hybrid framework that merges spiking neural networks (SNNs) with traditional artificial neural network (ANN)-based transformers to advance wireless foundation models (WFMs). Motivated by the human brain’s capacity for energy-efficient and noise-resilient information processing, SpikeWFM is designed to bolster WFM robustness against interference while preserving strong generalization across varied wireless contexts. Building on the achievements of large language models, WFMs utilize self-supervised pre-training on extensive datasets from multiple wireless environments to derive a unified embedding. This embedding facilitates a broad spectrum of downstream applications, such as channel estimation, positioning, beam prediction, and channel prediction. These models generally surpass task-specific architectures, offering greater adaptability to unforeseen conditions. Nevertheless, current WFMs are still susceptible to the noise and interference inherent in real-world wireless systems. To overcome this challenge, we embed spiking neurons within the transformer-based WFM structure. A theoretical analysis is provided to illustrate how the SNN-ANN hybrid combats noise and interference via event-driven processing and temporal sparsity. Our experiments demonstrate that SpikeWFM consistently exceeds standard ANN-based WFMs in both pre-training convergence speed and channel prediction precision. Further findings related to communication and sensing tasks will be detailed in the comprehensive journal edition of this research.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC