devices, edge servers, and cloud machines integrated into CPS are depicted. FL is rst a means to aggregate across edges and devices, enabled by low-latency data transmis-sion to generalize predictive and forecasting ability to service enablers. Meanwhile, data processing and AI-based applications [5] on Edge AI involve processing with edge and client sides directly to physical space. 2.1Federated Learning 2.1.1Federated learning paradigmIn conventional machine learning approaches, data is oen gathered into a centralized database. en sample data is drawn to train machine learning models until convergence. In an ideal scenario where data can be trans-ferred without communication cost, high computational capacity is available for training. ere are non-issues of data privacy, a model trained on the whole corpus will capture the statistical distribution of general data and should provide the highest accuracy for the entire training dataset.In the FL paradigm, each IoT device or Edge server owns its private data yet wants to cooperatively train a machine learning model without sharing individual data. e method of FL was rst proposed by McMahan et. al. [3] in which each participant trains a local model on their own data before sending it to a global server for aggrega-tion. e aggregated models are sent back to local for the next training round in an iterative way. Formulary, as described in Fig. 3, there are a set of participants along with their private data set. In each train-ing round, a global model w is aggregated from many local models trained locally on local datasets. erefore, the objective of learning problem is to nd an optimal aggre-gate function such that . Note that local models and aggregated models are iteratively updated over time. In the federated learning paradigm, each (communi-cation) round of federation oen involves many training epochs at local clients. Conventionally, Federated Averaging Aggregation (FedAvg) [3] shows a simple yet eective method to ag-gregate local models. In FedAvg, local models are optimized by stochastic gradient descent (SGD) using the same hyper-parameters including learning rate and training epochs over all participants. 2.1.2Federated learning systems and challengesCommonly, FL systems can be categorized into two types by the federation scale: cross-silo and cross-device federated learning [6][7]. Cross-silo federated learning is collaborative training between organizations such as hos-pitals and banks. In this type, the number of participants is small. However, local data is known to be signicant, with high-performance servers to serve as local training tremendous capacity. In contrast, in cross-devices FL, local models are trained collaboratively on end-user devices. In this type, the num-ber of participants in active training rounds could reach to million, with small computational capacity. Participants in this type can be using mobile devices or IoT sensor de-vices. Instead of training large datasets, this type faces a high communication cost when many participants update their model to the central server for aggregation.In our current research, instead of training on devices, data from end-user devices can be gathered to closed edge AI. AI models are then trained on edge before sending to cloud servers for aggregation. Edge AI may have a small capacity compared with local servers as cross-silo FL. Edge ig. F3 Federated Learning Paradigm1794-4 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#185