In the design of wireless receivers, DNNs can be combined with traditional model-based receiver algorithms to realize modular hybrid model-based/data-driven architectures that can account for domain knowledge. Such architectures typically include multiple modules, each carrying out a different functionality. Conventionally trained DNN-based modules are known to produce poorly calibrated, typically overconfident, decisions. This implies that an incorrect decision may propagate through the architecture without any indication of its insufficient accuracy. To address this problem, we present a novel combination of Bayesian learning with hybrid model-based/data-driven architectures for wireless receiver design. The proposed methodology, referred to as modular model-based Bayesian learning, results in better calibrated modules, improving accuracy and calibration of the overall receiver. We demonstrate this approach for the recently proposed DeepSIC MIMO receiver, showing significant improvements with respect to the state-of-the-art learning methods.

翻译：在无线接收机设计中，深度神经网络可与传统模型导向接收算法结合，构建能够融入领域知识的模块化混合模型/数据驱动架构。此类架构通常包含多个功能各异的独立模块。现有研究表明，采用常规训练方法的深度神经网络模块会产生校准不良的决策结果，且普遍存在过度自信问题。这意味着当某个模块出现错误决策时，错误信息可能在缺乏准确性提示的情况下沿架构传播。针对这一难题，本文提出一种将贝叶斯学习与混合模型/数据驱动无线接收机架构相结合的新型方法。所提出的"基于模块的模型导向贝叶斯学习"方法能生成校准效果更优的模块，全面提升接收机整体的决策准确度与校准性能。我们以近期提出的DeepSIC MIMO接收机为验证案例，实验结果表明该方法相较于现有最优学习方法实现了显著性能提升。