We consider the problem of decentralized hypothesis testing under communication constraints in a topology where several peripheral nodes are arranged in tandem. Each node receives an observation and transmits a message to its successor, and the last node then decides which hypothesis is true. We assume that the observations at different nodes are, conditioned on the true hypothesis, independent and the channel between any two successive nodes is considered error-free but rate-constrained. We propose a cyclic numerical design algorithm for the design of nodes using a person-by-person methodology with the minimum expected error probability as a design criterion, where the number of communicated messages is not necessarily equal to the number of hypotheses. The number of peripheral nodes in the proposed method is in principle arbitrary and the information rate constraints are satisfied by quantizing the input of each node. The performance of the proposed method for different information rate constraints, in a binary hypothesis test, is compared to the optimum rate-one solution due to Swaszek and a method proposed by Cover, and it is shown numerically that increasing the channel rate can significantly enhance the performance of the tandem network. Simulation results for $M$-ary hypothesis tests also show that by increasing the channel rates the performance of the tandem network significantly improves.
Leveraging Channel Noise for Sampling and Privacy via Quantized Federated Langevin Monte Carlo
对于人工智能的工程应用,贝叶斯学习比标准频率论学习具有显著的优势,包括量化不确定性的能力。Langevin Monte Carlo(LMC)是一种有效的基于梯度的近似贝叶斯学习策略,其目的是从模型参数的后验分布中产生样本。先前的工作集中在通过模拟调制的多址无线信道上的LMC的分布式实现。相比之下,本文提出了量化联邦LMC(FLMC),它集成了一位随机量化的局部梯度与通道驱动的采样。随机驱动采样利用信道噪声来促进蒙特卡罗采样,同时也起到隐私机制的作用。模拟和数字实现的无线LMC比较作为一个功能的差分隐私(DP)的要求,揭示了后者在足够高的信噪比的优势。