In this paper, we take an information-theoretic approach to understand the robustness in wireless distributed learning. Upon measuring the difference in loss functions, we provide an upper bound of the performance deterioration due to imperfect wireless channels. Moreover, we characterize the transmission rate under task performance guarantees and propose the channel capacity gain resulting from the inherent robustness in wireless distributed learning. An efficient algorithm for approximating the derived upper bound is established for practical use. The effectiveness of our results is illustrated by the numerical simulations.

翻译：本文采用信息论方法研究无线分布式学习中的鲁棒性问题。通过度量损失函数差异，我们给出了由无线信道不完美引起的性能恶化上界。此外，我们在任务性能保证条件下刻画了传输速率，并提出了由无线分布式学习固有鲁棒性带来的信道容量增益。为便于实际应用，我们建立了一种高效算法来逼近所推导的上界。数值模拟结果验证了本文结论的有效性。