AI served in tiny smart homes or buildings. us, the number of participants in FL is also higher than those in cross-silo. Moreover, Edge AI and devices also can col-laboratively train local models by taking training respon-sibility on smaller parts of models that split from original local models. ere are some vital challenges that must be solved during the training distributed to various participants, with heterogeneous local data distribution and complex system conguration. e challenges include expensive communi-cation, system heterogeneity, statistical heterogeneity, and security and privacy concerns. In addition to the main challenges, the other obstacles of federated learning can be pointed out: incentive mechanisms for participants, per-sonalization and multi-task on local sites, and fairness among participants, to name a few. ose obstacles make FL more challenging when deploying into real-world ap-plications, especially with the emerging Edge AI in hetero-geneous IoT environments.2.2Edge AI and use case2.2.1Edge intelligenceBig data processing requires more powerful methods, i.e., AI technologies, to extract insights that lead to better decisions and strategic business moves. In the last decade, with the great success of Deep Neural Networks (DNNs) [8], it has been possible to learn profound data representa-tions and become the most popular machine learning ar-chitecture. Considering that AI is functionally necessary to analyze vast volumes of data and extract insights quickly, a strong need exists to integrate Edge Computing and AI, creating Edge Intelligence. Edge Intelligence is not a simple combi-nation of Edge Computing and AI. e topic of Edge Intelligence is enormous and highly complex, covering many concepts and technologies intertwined in a dicult attitude [9]. Figure 4a depicts vital conceptual studies from the perspective of building an intelligent wireless network on Edge.2.2.2Artificial intelligence meets edge computingWe believe the fusion of AI and Edge Computing is natural and inevitable. ere is a close exchange and brings many outstanding advantages. On the one hand, AI pro-vides EC with technologies and methods that enable Edge Computing to unleash its potential and scalability with AI. In another aspect, EC provides AI with scenarios and platforms, and AI can expand its applicability with EC. e current research focuses on AI on Edge, including architecture, implementation, and application. AI on Edge is understood as deploying AI models on Edge. Edge AI for operating training and inference AI models combined with a device-cloud synergy. at aims to extract insights from large and scattered edges data with the satisfaction of algorithm performance, cost, privacy, reliability, eciency, etc. erefore, it can be interpreted as Edge AI. Fig. 4b. illustrates the dierences, complexity of AI, and data avail-ability across mobile, edge servers, and the cloud. e fact that data exists at devices such as edge clients, edge servers, and cloud is inversely proportional to its processing. As aforementioned, many solutions share data processing us-ing methods such as split learning, a combination of split learning (SL), and FL [9][10]. e primary purpose is to solve the computation load for both edge clients and not need to share data for the server.ig. F4 Artificial intelligence on Edge(a)(b)180 情報通信研究機構研究報告 Vol.68 No.2 (2022)4 スマートデータ利活用基盤技術
元のページ ../index.html#186