Research and Development on FL and Edge AI Over recent years, along with research direction in cross-data analysis, building big data platforms, and going toward the era of CPS with B5G. We deeply study and propose global and edge federation methodologies. e rst deals with non-IID spatial data distribution, while the latter assumes the edge environment consists of heteroge-neous IoT devices. is research adapts FL and edge AI to applications, including environmental monitoring systems and smart-home applications.3.1Cooperative model training among edge AI serversis research showed a FL system that overcomes heterogeneous IoT data issues among dierent edge sides of geographical areas [11]. In this system, edge AI emerged from training privately using local IoT data while aggregat-ing with global model training on a public dataset. Leveraging CRNN studies on air pollution prediction [12], we contribute a study to design and develop a federated learning framework on air pollution prediction with an illustration of high-ranking oxidant warning prediction. In applying pollution forecasting, data is collected from environmental stations in the Kanto region. Each station has a wide range of IoT sensors that measure environmen-tal indicators. If using a centralized machine learning ap-proach, all sensing data must be transferred to the central server for model training. Figure 5 shows the emerging edge computing environment in the system. Each prefec-ture is supposed to have its computational capacity with edge AI. Local models are trained on edge servers of the prefectures using environmental data. We also use public data to train the regional cloud model to enhance the ac-curacy of the master model while collaborating with models trained on local data. e global utilized training on public data experiments will be discussed aerward.a. MethodologyOn the local site, we employ the CRNN model [12] but the neural network structure is splitted into localized and (common) generalized parts to represent local distribution of monitoring station and sharing part across participants, respectively (Fig. 6a). Formally, ,,,, where ,,, are common and localized parts of participant k, respectively. ere is also the same predictive CRNN model, but it is trained on the global site with a public dataset. To aggregate on the global site, we adapt the average aggregation method [12] but with only common shared parts among participants are contribute to global update, such that ,,..., in each communica-tion round (Fig. 6b).b. Research outcomee experimental result showed that local models are trained separately on dierent spatial distribution data. However, once aggregated, the delivered model can perform similarly to the centralized model. Furthermore, by using public data for initializing and frequently updating the global model, the performance of the whole region remains transferable to local areas, even to a new area that has not participated in any communication round. 3ig. F5 Spatially distributed federated learning of regional pollution prediction system methodology1814-4 パブリック・プライベートデータを活用した予測モデリングを実現する連合学習によるエッジAI
元のページ ../index.html#187