Explainable Machine learning models for AI native radio access technologies (XAI-RAT)

机器学习（ML）已开始广泛用于增强5G无线电接入网络内各种组件的实现。此外，6G有望实现更高水平的自主性，改善人机交互，并在更多样化的环境中实现深度连接。因此，我们将接受这样一种愿景，即6G可能需要以ML可以修改物理（PHY）和介质访问控制（MAC）层的方式进行设计。在这样的人工智能（AI）原生无线电接入技术（RAT）中，ML可以例如使得能够学习个性化波形、调制方案、导频序列和代码，这不仅更有效地使用频谱，而且还最佳地适应于计算资源、收发器硬件和信道的实际限制。人工智能在细胞领域的应用虽然很有前途，但仍处于婴儿阶段，仍有重大挑战有待克服。 实现5G和6G之后支持AI的RAT愿景的关键挑战是：（i）训练数据的开销和可用性以及泛化的不确定性仍然是严重的开放性问题，（ii）缺乏可解释性，这源于人工智能工具通常被视为黑箱，因为很难开发分析模型来测试其正确性或解释其行为，以简单的方式，和（iii）部署问题，预见到缺乏互操作性和节能的硬件实现。 因此，当前研究计划的长期目标是开发具有可解释模型的AI原生无线电接入技术。其贡献在于：（i）提出低复杂度的可解释元学习技术，以解决实时在线培训问题，并大大提高支持AI的RAT的可信度，以及（ii）提出基于AI的近似计算（AC）方法，通过交易精度与能耗来实现积极的节能实施。关键思想还在于应用AI使AC框架与输入无关，并更好地利用错误补偿机制来提高AC性能。因此，主要成果是一种新的方法来研究AI原生RAT设计与其基于AI的节能实现之间的协同作用，其中在设计和实现过程中始终进行可解释性。因此，这项研究计划将有助于提高支持AI的RAT的可信度，加快工业采用，并支持标准化机构为集成AI模型提供关键见解。将优先考虑在机构/工业伙伴关系中实现相关项目。该计划的实现也是基于已经在期刊和会议上发表的研究成果。

Machine learning (ML) has started to be extensively used to enhance the implementation of various components within the 5G radio access network. In addition, 6G is expected to enable greater levels of autonomy, improve human machines interfacing, and achieve deep connectivity in more diverse environments. As such we are by to embrace a vision where 6G may need to be designed in a way that ML can modify parts of the physical (PHY) and medium access control (MAC) layers. In such an artificial intelligent (AI)-native radio access technologies (RATs), ML can for instance enable the learning of personalized waveforms, modulation schemes, pilot sequences and codes, which not only make a more efficient use of the spectrum but are also optimally adapted to practical limitations of the computational resources, the transceiver hardware and channel. The application of AI within the cellular domain, while promising, is still at its infant stages where significant challenges remain to be overcome. The key challenges for realizing the vision of AI-enabled RATs for beyond 5G and 6G are: (i) The overhead and availability of training data and the uncertainty in generalization that are still serious open issues, (ii) the lack of explainability that stems from AI tools being often treated as black boxes as it is hard to develop analytical models to either test their correctness, or explain their behaviours, in a simple manner, and (iii) deployment concerns that are foreseen from both the lack of interoperability and energy efficient hardware implementation. As such, the long-term goal of the current research program is to develop AI-native radio access technologies with explainable models. The contributions reside (i) in proposing low complexity explainable meta-learning techniques to address real-time on-line training issues and to considerably improve the trustworthiness of AI-enabled RATs, and (ii) in proposing AI-based approximate computing (AC) methodology for an aggressive energy efficient implementation by trading accuracy with energy consumption. The key idea is also on applying AI in making the AC framework input-independent and making better use of error compensation mechanisms to improve the AC performance. Therefore, the main outcome is a novel methodology to investigate the synergy between AI-native RATs design and its AI-based energy efficient implementation where explainability is carried out all along the design and implementation processes. As such, this research program will contribute to improving the trustworthiness of an AI-enabled RAT, speed up industrial adoption and support the standardization bodies toward providing key insights for integrating AI models. The realization of the related projects in institutional/industrial partnership will be privileged. The realization of this program is also based on research results already published in journals and conferences.