Comparison of Convergence Rates in Federated Learning and Federated Multi-Task Learning Using the CIFAR-10 Dataset

Many smart or cell phones have built-in distance, signal, and air pollution sensors. While collecting information, an acceleration registering device is a three-dimensional one and it can be applied in the gait analysis to address issues such as falls and health status determination. Indeed, the data is abundance in terms of quantity and some of the data may be of great concern in terms of privacy. In the time of Industry 4.0 the data has emerged as a key resource. Personal information/identity must not be maintained and hence cannot be stored at one place or all collected in a single place. AI models are moving to decentralized where a machine learning setting called Federated learning (FL) is being applied. FL has adversities such as statistical and systems heterogeneity. Actually, to better use shared information and build local models, Federated Multi-task learning (FMTL) has been devised. We also compare the number of iterations required to converge using CIFAR dataset of FL and FMTL. Several graphs illustrated in this paper show that convergence rates depend on the algorithm, number of communication rounds and number of clients or devices. Thus, it is clear that in some cases FL outperforms with FMTL in terms of convergence or conversely. However, it cannot be deduced that the type of FMTL always converges better than FL. The reliance on this graph is evident in this paper in order to as explain as prove the fact that, as the number of clients in FL rises, the rate of convergence declines. If ten communication rounds are employed with the use of the MOCHA algorithm, the model does not converge appropriately. The RMSE score declined from 1.14 to 1.02 throughout 20 epochs.