Efficient Integration of Distributed Learning Services in Next-Generation Wireless Networks

Distributed edge learning (DL) is considered a cornerstone of intelligence enablers, since it allows for collaborative training without the necessity for local clients to share raw data with other parties, thereby preserving privacy and security. Integrating DL into the 6G networks requires a coexistence design with existing services such as high-bandwidth (HB) traffic like eMBB. Current designs in the literature mainly focus on communication round-wise designs that assume a rigid resource allocation throughout each communication round (CR). However, rigid resource allocation within a CR is a highly inefficient and inaccurate representation of the system’s realistic behavior, especially when CR duration far exceeds the channel coherence time due to large model size or limited resources. This is due to the heterogeneous nature of the system, as clients inherently may need to access the network at different time instants. This work zooms into one arbitrary CR, and demonstrates the importance of considering a time-dependent design for sharing the resource pool with HB traffic. We first formulate a time-slot-wise optimization problem to minimize the consumed time by DL within the CR while constrained by a DL energy budget. Due to its intractability, a session-based optimization problem is formulated assuming a CR lasts less than a large-scale coherence time. Some scheduling properties of such multi-server joint communication scheduling and resource allocation framework have been established. An iterative algorithm has been designed to solve such non-convex and non-block-separable-constrained problems. Simulation results confirm the importance of the efficient and accurate integration design proposed in this work.

The proliferation of edge devices and advancements in Internet of Things technology have created a vast array of distributed data sources, necessitating machine learning models that can operate closer to the point of data generation. Traditional centralized machine learning approaches often struggle with real-time big data applications, such as climate prediction and traffic simulation, owing to high communication costs. In this study, we introduce random node entropy pairing, a novel distributed learning method for artificial neural networks tailored to distributed computing environments. This method reduces communication overhead and addresses data imbalance by enabling nodes to exchange only the weights of their local models with randomly selected peers during each communication round, rather than sharing the entire dataset. Our findings indicate that this approach significantly reduces communication costs while maintaining accuracy, even when learning from non-IID local data. Furthermore, we explored additional learning strategies that enhance system accuracy by leveraging the characteristics of this method. The results demonstrate that random node entropy pairing is an effective and efficient solution for distributed learning in environments where communication costs and data distribution present significant challenges.

Communication-efficient distributed cubic Newton with compressed lazy Hessian

Distributed swarm learning for communication-efficient capsule endoscopy classification on solving Non-i.i.d data

Analyzing massive amounts of data using complex machine learning models requires significant computational resources. The conventional approach to such problems involves centralizing training data and inference processes in the cloud, i.e., in data centers. However, with the proliferation of mobile devices and increasing application of the Internet-of-Things (IoT) paradigm, very large amounts of data are collected at the edges of wireless networks, and due to privacy constraints and limited communication resources, it is undesirable or impractical to upload this data from mobile devices to the cloud for centralized learning. This problem can be solved by distributed learning at the network edge, by which edge devices collaboratively train a shared learning model using real-time mobile data. The avoidance of raw-data uploading not only helps to preserve privacy but may also alleviate network-traffic congestion and minimize latency. With that said, distributed training still requires a substantial amount of information exchange between devices and edge servers over wireless links. In the process, wireless impairments such as noise, interference, and imperfect knowledge of channel states can significantly slow down distributed learning (e.g., convergence speed) and degrades its performance (e.g., learning accuracy). This makes it crucial to optimize wireless network performance so as to support the efficient deployment of distributed learning algorithms. On the other hand, distributed learning algorithms provide a powerful tool-set for solving complex problems in wireless communication and networking. One important framework, called federated learning (FL), enables users to collaboratively learn a shared model while helping to preserve local data privacy. The application of FL can endow edge devices with capabilities of user behavior prediction, user identification, and wireless environment analysis. As another example, distributed reinforcement learning is capable of leveraging distributed computation power and data to solve complex optimization and control problems that arise in various use cases, such as network control, user clustering, resource management, and interference alignment. To cover this paradigm of distributed learning over wireless networks, this two-part Special Issue features papers dealing with two main research challenges: a) optimization of wireless network performance for efficient implementation of distributed learning in wireless networks, and b) distributed learning for solving communication problems and optimizing network performance.

In this paper, we consider communication-efficient over-the-air federated learning (FL), where multiple edge devices with non-independent and identically distributed datasets perform multiple local iterations in each communication round and then concurrently transmit their updated gradients to an edge server over the same radio channel for global model aggregation using over-the-air computation (AirComp). We derive the upper bound of the time-average norm of the gradients to characterize the convergence of AirComp-assisted FL, which reveals the impact of the model aggregation errors accumulated over all communication rounds on convergence. Based on the convergence analysis, we formulate an optimization problem to minimize the upper bound to enhance the learning performance, followed by proposing an alternating optimization algorithm to facilitate the transceiver design for AirComp-assisted FL. As the alternating optimization algorithm suffers from high computation complexity, we further develop a knowledge-guided learning algorithm that exploits the structure of the analytic expression of the transmit power to achieve computation-efficient transceiver design. Simulation results demonstrate that the proposed knowledge-guided learning algorithm achieves a comparable performance as the alternating optimization algorithm, but with a much lower computation complexity. Moreover, both proposed algorithms outperform the baseline methods in terms of convergence speed and test accuracy.

Federated learning (FL) in mobile edge computing has emerged as a promising machine-learning paradigm in the Internet of Things, enabling distributed training without exposing private data. It allows multiple mobile devices (MDs) to collaboratively create a global model. FL not only addresses the issue of private data exposure but also alleviates the burden on a centralized server, which is common in conventional centralized learning. However, a critical issue in FL is the imposed computing for local training on multiple MDs, which often have limited computing capabilities. This limitation poses a challenge for MDs to actively contribute to the training process. To tackle this problem, this paper proposes an adaptive dataset management (ADM) scheme, aiming to reduce the burden of local training on MDs. Through an empirical study on the influence of dataset size on accuracy improvement over communication rounds, we confirm that the amount of dataset has a reduced impact on accuracy gain. Based on this finding, we introduce a discount factor that represents the reduced impact of the size of the dataset on the accuracy gain over communication rounds. To address the ADM problem, which involves determining how much the dataset should be reduced over classes while considering both the proposed discounting factor and Kullback-Leibler divergence (KLD), a theoretical framework is presented. The ADM problem is a non-convex optimization problem. To solve it, we propose a greedy-based heuristic algorithm that determines a suboptimal solution with low complexity. Simulation results demonstrate that our proposed scheme effectively alleviates the training burden on MDs while maintaining acceptable training accuracy.

Energy management and demand response with intelligent learning for multi-thermal-zone buildings

Federated learning is a distributed learning framework designed to protect user privacy, widely applied across various domains. However, existing federated learning algorithms face challenges, including slow convergence, significant loss fluctuations during aggregation, and imbalanced client sampling. To address these issues, this paper introduces a high-performance federated learning aggregation algorithm. This algorithm combines a cyclic adaptive learning rate adjustment strategy with client-weighted random sampling, addressing the aforementioned problems. Weighted random sampling assigns client weights based on their sampling frequency, balancing client sampling rates and contributions to enhance model aggregation. Additionally, it adapts the learning rate based on client loss variations and communication rounds, accelerating model convergence and reducing communication costs. To evaluate this high-performance algorithm, experiments are conducted using well-known datasets MNIST and CIFAR-10. The results demonstrate significant improvements in convergence speed and loss stability. Compared to traditional federated learning algorithms, our approach achieves faster and more stable convergence while effectively reducing training costs.

Federated and fusion learning methods are state-of-the-art distributed learning approaches which enable model training without collecting private data from users. While federated learning involves lower computation cost as compared to fusion learning, the overall communication cost is higher due to a large number of communication rounds between the clients and the server. On the other hand, fusion learning reduces the overall communication cost by sending distributions of features and model parameters using only one communication round but suffers from high computation cost as it needs to find the distributions of features at the client. This paper presents hybrid fusion learning, a system that leverages hierarchical client-edge-cloud architecture and builds a deep learning model by integrating both fusion and federated learning methods. Our proposed approach uses fusion learning between the client and the edge layer to minimise the communication cost whereas it uses federated learning between the edge and the cloud layer to minimise the computation cost. Our results show that the proposed hybrid fusion learning can significantly reduce the total time taken to train the model with a small drop of around 2% in accuracies as compared to the other two algorithms. Specifically, our results show that fusion and federated learning algorithms take up to 26.28% and 9.74% higher average total time to build the model, respectively, than the proposed hybrid fusion learning approach.

The next-generation wireless network is expected to be highly dense with a large number of Device-to-Device (D2D) communication enabled IoT devices in fog computing based cellular networks. The dense deployment of the heterogeneous network architecture is expected to fulfill the smart devices’ growing data demand, lower power consumption, and lower latency constraint. The 5G technology is expected to have such multi-tier architecture with various computational capability and radio enabled IoT devices. However, the coexistence of such heterogeneous network model spawns research challenges such as interference management, non-uniform computational capacity with non-uniform devices connectivity and service deadline. Thus, in this paper, we propose a coexistence of D2D-IoT (D-IoT) and fog computing model in cellular networks and formulate the resource allocation problem in such a multi-tier architecture considering different kinds of interference, data rate, and latency altogether as an optimization problem and further propose a distributed many-to-many stable matching based solution. Through extensive theoretical and simulation analysis, we have shown the effect of different parameters on the resource allocation objectives and achieve more than 94% of optimum network performance.

Radio resource management for user-relay assisted OFDMA-based wireless networks

Over-the-air federated learning (AirFL) has developed as a communication-efficient solution to enable distributed machine learning over edge devices by integrating computation and communication into a joint design. However, the aggregation error caused by noise and wireless channel fading may lead to a compromised learning performance in an AirFL system. Potential malicious participants can also cause this compromise such that decoding the individual uploaded model updates at the server side to perform security examination is still imperative for AirFL. In this paper, we investigate a novel multi-objective resource allocation problem to combat against such aggregation error and improve the information transmission efficiency of uplink by jointly optimizing the device selection, power allocation and receive scalar control. The objective of the multi-objective optimization problem (MOP) is to minimize the average mean squared error (MSE) of the over-the-air aggregation for different communication rounds and maximize the long-term energy efficiency (EE) of the system. Considering the complexity and dynamic environment, we present a deep reinforcement learning (DRL) based framework to solve the MOP. The MOP is firstly decomposed into multiple subproblems. Each subproblem is modelled as a neural network, which can be addressed by the proposed learning-based resource allocation (LRA) algorithm. The numerical results illustrate that our proposed approach can effectively tackle the MOP and outperform the benchmark approaches with appropriate training mechanism and reward design.

The limited communication and computing resources, as well as the rising concerns about the privacy protection, bring significant challenges to the massive data training and analysis in vehicular networks. To address these challenges, in this paper a platoon-based distributed learning framework design for data learning is carried out, where the vacant computation resources of vehicle platooning networks are leveraged. In the proposed framework, a 2-phase Markovian stochastic process is utilized to depict the learning service heterogeneity for each participating vehicle. Meanwhile, we propose a joint scheduling and resource allocation scheme for efficiency-oriented distributed learning to maximize the learning accuracy subject to a given learning time constraint. The optimization problem is solved as follows. First, given the scheduled vehicles, the communication resource allocation is modeled as a minimum-maximum problem to minimize the learning delay of each learning round. Subsequently, an efficiency-oriented unbiased global aggregation policy is proposed to explore the convergence difference between partial scheduling and total scheduling. Considering the learning convergence and remaining time, an on-demand scheduling scheme is introduced to determine the number of scheduled vehicles. Finally, combining the learning efficiency of each vehicle with the scheduled number of vehicles, the scheduled vehicle set is selected. Simulations results show that the proposed scheduling policy can schedule the number of participating vehicles on demand based on the trade-off between learning performance and learning latency.

We propose a distributed learning algorithm for the resource allocation problem in Device-to-Device (D2D) wireless networks that takes into account the throughput estimation noise. We first formulate a stochastic optimization problem with the objective of maximizing the generalized alpha-fair function of the network. In order to solve it distributively, we then define and use the framework of noisy-potential games. In this context, we propose a Binary Log-linear Learning Algorithm (BLLA), which is distributed across cells and converges to a Nash equilibrium of the resource allocation game. This equilibrium is also an optimal for the resource allocation optimization problem. A key enabler for the analysis of the convergence are the proposed rules for computation of resistance of trees of perturbed Markov chains. The convergence of BLLA is proved for bounded and unbounded noise, with fixed and decreasing temperature parameter. A sufficient number of estimation samples is also provided that guarantees the convergence to an optimal state in a single cell scenario and close to an optimal state in a multi-cell scenario. We assess the performance of BLLA by extensive simulations by considering both bounded and unbounded noise cases and show that BLLA achieves higher sum data rate compared to the state-of-the-art.