Foundation model (FM) has shown great potential in various downstream tasks, such as CLIP, a vision-language pretrained model that effectively aligns semantics in text and image spaces. Federated learning (FL), as a distributed training paradigm, can provide adequate data for adapting FM while overcoming challenges such as communication pressure and user privacy leakage. However, in edge computing with a large scale of edge devices, vanilla FL faces challenges, including limited computation resources, inadequate bandwidth, non-IID and multi granularity data caused by the diversity of devices and users. To feasibly and efficiently collaborate with the large and diverse edge devices in adapting FM, we propose a novel personalized and multi-granularity FL framework (PMG-FL) that provides a personalized lightweight prompt for each edge device while considering interactions among same-granularity edge devices and cross-granularity edge devices. In particular, we introduce prompt training to adapt FM locally, which can mitigate computation and communication pressure with few learnable prompt parameters. Based on the prompt parameters, we design a distance-based prompt aggregation mechanism to capture similarities among same-granularity edge devices and aggregate the personalized prompt for each edge device, addressing the challenges posed by non-IID data. Furthermore, we design a cross-granularity guidance mechanism that leverages the correlation of semantic knowledge among edge devices with multi-granularity data. Extensive experimental results demonstrate the superiority of PMG-FL over the alternative approaches, with robust performance on both IID and non-IID data.

In this paper, we propose GenG, a generic time series data generation approach for edge intelligence that incorporates knowledge from different domains to synthesize high-fidelity and controllable time series data resembling to different IoT devices. Specifically, GenG decomposes the time series data generation task into two subtasks, the first subtask is to finetune a Large Language Model (LLM) in a self-training method to harness its outstanding knowledge and reasoning capacities for explainable data generation, solving the problem of what to generate. The second one focuses on generating high-quality and controllable time series data conditioning on the output of the finetuned LLM, solving the problem of how to generate. Furthermore, a two-stage generation process is proposed to increase the quality of the generation results by introducing both the abstract and detailed guidance signals, which also enables flexible control over the generation results and ensures synthesized data with consistent features. During deployment, GenG can be arranged in a cloud-edge collaboration way, where the cumbersome LLM and light-weight generation model are placed on the cloud and edge, respectively, fitting well with the resource-constrained edge intelligence. Experimental results in different generation tasks demonstrate GenG's efficiency in reasoning about the generation task and synthesizing high-fidelity time series data with controllable features.

Federated learning has a communication bottleneck since a considerable number of parameters are transferred between the central server and edge devices. Therefore, some prior works proposed compression methods by using the top-k sparsification to solve communication-efficient issue. However, we observe that this method affects the accuracy in the early stage of training. To address the problem, we propose a novel parameter upload mechanism, viz. FedBF16-Dynamic. In the early communication rounds of training, we apply the brain floating-point (Bfloat16) for transmission numerical type to upload model parameters, which reduces communication cost comprehensively. In subsequent communication rounds, there are two upload schemes for the edge devices with different levels of uplink bandwidth. Compared with the baseline, simulation results show that the proposed FedBF16-Dynamic scheme reduces communication cost and achieves higher performance within the least amount of time in various network environments.

Split learning (SL) is a promising approach to enable edge devices for training artificial intelligence (AI) models. We consider the heterogeneous devices collaborating with a base station (BS) to train an AI model in a distributed manner, for privacy concerns and communication resource constraints. However, due to device heterogeneity and channel uncertainty, the differentiated computing capabilities among devices and the time-varying connections between the BS and devices affect the resource utilization of the BS and devices, and thus, the training delay. In this paper, we design an adaptive split learning (ASL) scheme that determines split point selection and computing resource allocation. We formulate an optimization problem to minimize the average training latency subject to a long-term energy consumption constraint. The major challenges in solving this problem are the lack of future information and mixed integer programming (MIP). To solve it, we propose an online algorithm, named OPEN. Leveraging the Lyapunov theory, the proposed algorithm transforms the problem into a new problem only with the current information, which is a MIP problem. Then, a two-layer optimization method is proposed to solve the MIP problem. Extensive simulation results demonstrate that the ASL scheme can reduce the average training delay and energy consumption by 53.7% and 22.1%, respectively, as compared to the existing SL schemes.

With the growing demand for various applications in intelligent rail transit, the burden of information transmission is aggravated. Meanwhile, high-mobility trains, time-varying channels and limited package transmission delays bring great challenges to the quality of packet transmission in train-to-train (T2T) communications. In this paper, to guarantee the transmission quality and reduce the deployment cost, wireless ad hoc network as a novel technology is applied to the communications-based train control (CBTC) system. In order to further improve the transmission efficiency of multi-hop ad hoc networks, a federated soft actor-critic (FeSAC) approach is proposed for joint optimization of relay selection, subchannel allocation and power control. The goal of the FeSAC algorithm training is to make the throughput of T2T links as large as possible with less energy consumption, while ensuring the link reliability and queuing delay requirements of T2T communications. To synthesize the training results of each agents and accelerate model convergence, the FeSAC algorithm aggregates the network parameters by calculating the weighted mean value based on the reward of individual edge intelligent agents. The training tasks are offloaded to edge intelligent nodes, which makes parameter aggregation more easily through interconnection. Evaluated by the simulation, the proposed FeSAC algorithm is more capable of increasing throughput and reducing energy consumption compared to other algorithms.

Artificial Intelligence-Generated Content(AIGC) represents a technique that AI automatically generates various forms of content that meet the personalized requirements of humans. To offer an easy access to AIGC techniques, an edge computing system is introduced with heterogeneous computing capability to process users' text-to-image AIGC requests. For text-to-image AIGC tasks, inference steps configuration of diffusion models has a great impact on inference time and quality of generated results. Thus, it is necessary to design an online scheduling scheme to dispatch users' AIGC requests to suitable servers and tune proper inference steps for AIGC requests in order to achieve better quality of service. Based on this motivation, we define a novel online text-to-image AIGC request scheduling problem under a heterogeneous edge computing scenario with the objective of striking a balance between tardiness of AIGC requests and quality of generated results. We formulate it into an integer programming problem. Then we transform this problem into a Markov Decision Process(MDP) model and adopt a deep reinforcement learning-based algorithm to solve this online problem. Simulation results show that our approach performs better and has a strong generalization ability, compared with the baseline random policy and greedy policy.

As we know, the training of Large Language Models (LLM) is time-consuming and expensive. Its training efficiency is often affected by both the heterogeneous computing devices and heterogeneous communication networks in the computing cluster. In recent years, new computing devices and technologies such as NVIDIA H200 and Compute Express Link (CXL) 3.0 have been proposed, bringing new opportunities for improving the training efficiency of LLM. However, the actual deployment difficulty and cost of these new devices or technologies are extremely high, so it is difficult for researchers to evaluate their impacts or improvements on LLM training. In order to solve this problem, this paper introduces a simulation tool named HeterSim, and proposes to simulate and evaluate LLM training in CXL-based heterogeneous computing clusters using HeterSim. This article takes the LLM called LLaMA as a simulation example, and successfully simulates and analyzes the impact of heterogeneous computing and CXL technologies on LLM training. We hope that this article can provide researchers with new ideas for simulating and analyzing LLM training, and help researchers explore the impact of emerging technologies on LLM training at low cost.

Network traffic prediction is critical in wireless network management by allowing a good estimate of the traffic trend, which is also an important approach for detecting traffic anomalies in order to enhance network security. Deep-learning-based method has been widely adopted to predict network traffic matrix (TM) though with the main drawbacks in high complexity and low efficiency. In this paper, we propose a traffic prediction model based on Particle Swarm Optimization (PSO) and LightGBM (PSO-LightGBMTM), which optimizes the LightGBM parameters for each network flow by PSO so that LightGBM can adapt to each of the network traffic flow. Compared with existing commonly used deep learning models, our model has a more straightforward structure and yet outperforms existing deep learning models. Sufficient comparison tests on three real network traffic datasets, Abilene, GÉANT, and CERNET have been conducted, and the results show that our model provides more accurate results and higher prediction efficiency.

To broaden their domain knowledge coverage, large language models (LLMs) increasingly incorporate extensive corpus data from various industries. These heterogeneous datasets are often maintained by different stakeholders, where issues of data heterogeneity, privacy, and the network cost of data transmission have attracted much attention. To address these challenges, researchers have studied the integration of LLMs with knowledge graphs to manage data heterogeneity and with edge computing to ensure data privacy and transmission efficiency. In this work, we introduce a reputation system and a spot-check mechanism for a decentralized knowledge inference system in which edge nodes can collaborate with others for knowledge sharing while preserving their data privacy. We then use an evolutionary game model to study the dynamic decision-making between requestors and workers. In addition, we demonstrate that when the requestor can correctly identify a malicious worker with a probability greater than 0.5, the system evolves to maximize overall social welfare. Moreover, we show that higher reward values and higher model quality accelerate the maximization of social welfare.

Mobile Edge Computing (MEC) can effectively speed up data processing and improve Quality of Service (QoS) by offloading Mobile Users' (MUs') tasks to nearby Edge Servers (ESs). However, due to the individual rationality of entities (i.e., ESs and MUs) in MEC networks, they may be reluctant to participate in the computation offloading process without reasonable resource pricing or compensation. To address the challenge, we propose a Two-stage Stackelberg game-based computation Offloading and Resource Pricing mechanism (TORP). Specifically, we first introduce a broker between ESs and MUs, which rents computation resources from ESs and provides services to MUs. Next, we formulate the interactions among the broker, MUs, and ESs as a two-stage Stackelberg game, aiming to maximize their respective utilities. Meanwhile, we propose a Gradient-Ascent-Based Dynamic Iterative Search Algorithm (GADISA) to determine service subscription strategies of MUs and an Alternating Iteration-Based Resource Pricing and Task Offloading Algorithm (AIPOA) to determine resource sharing strategies of ESs. Simulations show that TORP greatly outperforms other benchmarks in improving the utilities of three entities.