Federated learning is a distributed, and collaborative learning method that allows different edge devices with different datasets to work together to train a global model. In this learning, a single global model is stored in a centralized cloud infrastructure. At first, the global model is shared with devices with initial weights. Now the edge device collects the real-time data and trains the model locally with the new data for one or several iterations in order to update the weights so that the loss function is minimized [19]. The updated weights are sent to a centralized server. Here the data are not sent to the centralized server. Only the weights are sent to the server with encryption [4]. The centralized server receives the updated weights from several edge devices. It computes the average of updated weights, and then it updates the weights of the global model. Then the global model is again shared with edge devices. The concept of federated learning is shown in Figure 7. Federated learning could cater to the needs of modern IoT-based applications and turns out to be the basis for next-generation artificial intelligence [20].

As discussed above, in centralized learning, the server sends a model, which initially gets trained in the server, to edge devices. In an edge device, the model gets training with local data, and the updated model parameters are sent to the server. In this way, the server aggregates all the parameters and resends the updated model to devices. In the case of decentralized, federated learning such as Peer-to-Peer [22] or in a ring topology [23], in addition to training the model with local data, each device performs the updating of parameter values with a Gossip algorithm [24,25]. As multiple devices are involved in training the model simultaneously, the training time is reduced [26]. Moreover, in federated computing, privacy and security of data are maintained [27]. Federated learning is more appropriate for utilizing low-costing machine learning models on devices and sensors [28].

As far as the data in different clients are homogeneous in which the feature space of local data in participating clients is the same, updating the global model with updated weights of clients would be easy, and such data federated learning is called horizontal learning, and it leads to an effective global solution [29]. However, updating the global models becomes a challenge when the data in devices are of the kind non-Independent and Identically Distributed (non-IID) which makes the convergence of the global model becomes difficult [30]. This issue can be resolved by developing a personalized local model for each client [31], and then the personalized local models can be merged into a shared global model with the help of the Bayesian fusion rule, as discussed in [32,33]. In another research work [34], a branch-wise averaging-based aggregation method has been proposed, which guarantees convergence of the global model. In another work [35], feature-oriented regulation method has been proposed to establish a firm structure information alignment across collaborative models.

Federated learning is of two kinds, centralized [21] and decentralized (Figure 8).

Edge devices such as IoT, webcam, robots, intelligent medical equipment, etc., are very useful for many healthcare applications during a pandemic, say, for example, COVID-19. Both shortages of reliable datasets, limited hardware, and power support of edge devices prohibit the usage of deep learning models in them. However, in Deep Transfer Learning (DTL), the knowledge of an already learned model is used to solve a new task, as in Figure 9. DTL significantly reduces training time and the requirement of resources for a target domain-specific task for a fixed feature extraction or fine-tuning [36,37].

In contrast to conventional machine learning, where learning takes place in isolation, transfer learning uses knowledge learned from other existing domains while learning for a new task. Transfer learning is of two types, homogeneous and heterogeneous. In homogeneous transfer learning, the source and target domains are the same. It means that the feature space is the same, whereas, in heterogeneous transfer learning, the feature space is not the same. Different methods, namely, instance-based methods, feature-based methods, parameter-based methods, and relation-based methods, are being used for homogeneous transfer learning, whereas for heterogeneous domains, only feature-based methods are being used. There are two types of methods, namely asymmetric feature transformation and symmetric in heterogeneous transfer learning. In asymmetric transfer learning, one of the domains is transformed into the other, and this method is found to be effective when both domains share the same label space. In symmetric transfer leaning, both domains are transformed into a common feature space.

Large machine learning models have millions of parameters associated with them, which makes the deployment of the model in edge devices infeasible. So, the knowledge gained from large models is transferred to small models which run on edge devices. Here, the large models serve as the teacher model, and the small model is such as the student model. The teacher model refers to a larger model, and it alone gets pre-trained. The learned knowledge from the teacher model is transferred to the student model through knowledge distillation, as in Figure 10. The knowledge distillation helps in improving the accuracy of the student model despite the constrained hardware [38]. Different knowledge distillation techniques, namely response-based distillation, where the prediction performance of output layers of the teacher model and student model are compared using a loss function which shows the difference between the models and the loss function is minimized so that the accuracy of student model approaches that of the teacher model. In contrast to response-based distillation, in feature-based distillation teacher model distills the intermediate features of the student model. Here, the position of distillation is moved prior to the output layer [39,40]. Here the student mimics by minimizing the loss function that is computed according to the intermediate layers. Relation based distillation is one in which the difference between the relationship between different feature maps is captured as a gram matrix, and the corresponding loss function is minimized. The student model reduces the computation cost and memory usage [41].

In addition, knowledge distillation is also used to tackle the problems, namely, heterogeneity in the local data of different devices and heterogeneity in the architecture of the models in edge devices. Federated learning can be implemented through knowledge distillation, as discussed in [42,43].