Open Research Direction to Edge AI Collaboration on Federated LearningIn a conventional AI system, triggers take place in the cloud to provide customer service. However, in this distrib-uted system, we assume that customers also contribute to training AI models using their own local data. To illustrate, local authorities can use weather and atmospheric data related to congestion [15]. Transport companies can use the cameras and sensors of trucks and taxis inside to predict trac events, for example, using the MM [16] traf-c event model. Additionally, we’re considering using public data on our platform to train models in the cloud to generalize AI models to participants. However, the above studies bring a lot of points and benets to the commu-nity. However, the AI model still uses distributed ML models, making it challenging to deploy in large spaces such as cities, transportation companies, taxis, etc. Another solution that opens a new direction is FL applied to solve the above diculty. However, FL oers many advantages in distributed ML models for customers with compact design congurations. And the FL faces the challenge of selecting such models as proximity, ensuring data collec-tion from dierent sensors. Because clients have congu-rable structures (e.g., processor, memory, computer, power, etc.), it’s hard to respond in constant time. So, to overcome this diculty and challenge, we are developing the solution using federated & split learning for the EDGE AI applica-tion. We hope our solution can overcome the problem in FL and still ensure computing power, data processing, and security in sharing private data.ConclusionFrom the viewpoint of data privacy, federated learning provides a mechanism to cooperatively train machine learning models while preventing any information leakage. On the other hand, federated learning challenges many deployment obstacles in real application with heteroge-neous participant data distribution and information and communication technology infrastructure.In our recent research, we deeply discover the frontier of using edge computing by delivering Edge AI in the context of federate learning paradigm along with applica-tion in smart-home including sleep quality prediction and health monitoring. We also tackle the problem of non-IID distribution by introducing public-private data cooperation and the newly adopted cross-silo federation on spatial skewness when providing environmental pollution predic-tion services on a large region that consist of many local areas.In the era of cyber-physical integration along with the advanced in telecommunication, e.g., 5G / 6G, we address the challenges and research direction on federated learning with edge computing by emerging Edge AI for training and reasoning while delivering robust aggregation algorithms dealing with various types of data distribution skewness. In the context of Beyond 5G / 6G adaptation, espe-cially on the Edge AI behavior support study [1] whereas the advantage of Edge AI as the combines of edge comput-ing and AI to train ML models and inference in high-ca-pacity edge server, low-latency data transferring and heterogeneous IoT data, we open the research to leverage Edge AI throughout the participants on Federated learning paradigms.In such a system, we address research obstacles that adopt Beyond 5G / 6G requirements in both Federated learning and Edge AI consideration. Firstly, Edge AI needs to optimize machine learning processing within the edge environment including network conditions and resources available. For that purpose, it must be possible to adjust the size of the respective ML models. is can be done by shrinking them with techniques such as Split Neural Network [17] on edge environments. To illustrate, small computing in smart cars or trucks cannot serve to train models but they can send IoT data and share resources with Edge AI by partially training models leveraging split neural network techniques.Secondly, it will also be necessary to develop federated learning technology that combines supper-diverse data streams and heterogenous and non-IID data distribution [18], for example dierent data IoT data capturing between transportation companies and mobility service providers, to improve performance of prediction outcomes. To over-come these obstacles, in the near future, we are conducting aggregation methods that deal with skewness in data fea-tures, spatial-temporal sampling by leveraging individual distribution from edge sites.eferencesR1NICT Beyond 5G / 6G white paper version 2.0 (released in June 2022), https://beyond5g.nict.go.jp/images/download/NICT_B5G6G_WhitePaperEN_v2_0.pdf2Khaled B. Letaief, Yuanming Shi, Jianmin Lu, and Jianhua Lu, “Edge Artificial Intelligence for 6G: Vision, Enabling Technologies, and Applications,” IEEE Journal on Selected Areas in Communications, vol.40, issue 1, pp.5–36, Jan. 2022.45184 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#190