Chuan Sun

Large language models (LLMs) have demonstrated strong capability across diverse applications. Local deployment of LLMs offers significant benefits such as enhanced privacy and personalization. Nevertheless, performing inference with LLMs incurs substantial computational costs, making it challenging to deploy LLMs on consumer devices like personal computers and smartphones. Network pruning is a promising approach to reduce the sizes and computational demands of LLMs without sacrificing functionality. However, existing pruning methods for LLMs often adopt a uniform pruning ratio across all Transformer layers, neglecting their varying contributions to performance, and potentially undermining the effectiveness of the pruned LLMs. In this paper, we propose the Fisher-based Layer-wise Adaptive Sparsification (FisherLAS) method to address this problem. It leverages the mean of the Fisher information matrix to determine the sensitivity of each Transformer layer and adaptively optimizes the allocation of the layer-wise pruning ratio. Theoretical analysis shows the rationality of its design. Extensive experiments on 7 open-source LLMs demonstrate the superiority of FisherLAS over 8 state-of-the-art methods, outperforming them on average by 26.86% and 4.81% respectively in terms of perplexity on WikiText2 and classification accuracy on various tasks.

Mobile edge computing (MEC) offers an advanced solution to offload tasks from user devices (UDs) to the network edge, alleviating network congestion and reducing task execution latency. Nevertheless, current computation offloading still faces two critical challenges: high transmission latency due to poor wireless channel quality, and task privacy leakage due to the potential for eavesdropping. Intelligent reflecting surface (IRS) is a promising technique to enhance channel quality and security in MEC networks. In this paper, we consider a multi-region IRS-assisted MEC network that jointly optimizes computation offloading and IRS phase configuration to minimize the task execution latency for UDs. The optimization problem is NP-hard and requires distributed decision-making in multiple regions. Therefore, we propose a multi-agent deep reinforcement learning algorithm with an attention mechanism and Gumbel-Softmax. Specifically, the attention mechanism helps dynamically select key information and improve decision-making efficiency, while Gumbel-Softmax is used to solve the problem of the non-differentiable discrete action spaces. blackSimulation results demonstrate that the proposed algorithm outperforms baseline schemes by reducing the task execution latency and improving the offloading rate by 27.2% and 60.7% on average in the considered MEC networks.

Federated Foundation Models (FedFMs) represent a distributed learning paradigm that fuses general competences of foundation models as well as privacy-preserving capabilities of federated learning. This combination allows the large foundation models and the small local domain models at the remote clients to learn from each other in a teacher-student learning setting. This paper provides a comprehensive summary of the ten challenging problems inherent in FedFMs, encompassing foundational theory, utilization of private data, continual learning, unlearning, Non-IID and graph data, bidirectional knowledge transfer, incentive mechanism design, game mechanism design, model watermarking, and efficiency. The ten challenging problems manifest in five pivotal aspects: “Foundational Theory,” which aims to establish a coherent and unifying theoretical framework for FedFMs. “Data,” addressing the difficulties in leveraging domain-specific knowledge from private data while maintaining privacy; “Heterogeneity,” examining variations in data, model, and computational resources across clients; “Security and Privacy,” focusing on defenses against malicious attacks and model theft; and “Efficiency,” highlighting the need for improvements in training, communication, and parameter efficiency. For each problem, we offer a clear mathematical definition on the objective function, analyze existing methods, and discuss the key challenges and potential solutions. This in-depth exploration aims to advance the theoretical foundations of FedFMs, guide practical implementations, and inspire future research to overcome these obstacles, thereby enabling the robust, efficient, and privacy-preserving FedFMs in various real-world applications.

Edge caching presents a promising avenue for mitigating backbone network congestion by strategically caching frequently accessed content at the network periphery. As most current edge caching solutions are designed for single-modal content requests, they cannot deal with the increasing volume of multi-modal content requests. In this study, we investigate the issue of multimodal content caching in edge-cloud networks. Firstly, we establish a heterogeneous edge-cloud network adept at caching multimodal content proximate to end-users to facilitate expeditious content delivery. By leveraging latent representations of multimodal content, we identify distinct user request modalities for multimodal content. Subsequently, we formulate caching replacement operations as a Markov Decision Process (MDP) aimed at minimizing user-content access latency. Moreover, we propose a decentralized multimodal content caching framework at the network edge based on federated deep reinforcement learning. This framework affords distributed decision-making and learning capabilities, thereby alleviating the strain on centralized resources and augmenting caching efficacy. To demonstrate the efficacy of our proposed framework, we conduct comprehensive experiments utilizing the Noah-Wukong dataset. Experimental results provide evidence that our framework reduces average latency by up to 10% compared to traditional methods, highlighting its proficiency in enhancing cache performance in edge-cloud networks.

With the advancements in AI-generated content (AIGC) technologies and edge cloud networks, the demand for AIGC services is increasing. The diffusion model stands out for its ability to generate images from text. However, its high computational requirements challenge user devices, leading to ongoing efforts to improve inferring speed while preserving image quality. In this study, we propose a novel edge-user collaborative inferring framework. By capitalizing on collaboration among the devices in edge and user, the proposed framework optimizes the service delay, network resource consumption, and user device computing resource consumption of the user’s AIGC service, and improves the user’s QoE. To demonstrate the efficiency of our proposed framework, we conduct comparative experiments and ablation experiments on a variety of datasets. Experimental results show that the proposed framework achieves better-generated image quality and reduces a large number of computing resources and network consumption, improving user QoE.

Mobile edge-cloud computing networks can provide distributed, hierarchical, and fine-grained resources, and have become a major goal for future high-performance computing networks. The key is how to jointly optimize service caching and computation offloading. However, the joint service caching and computation offloading problem faces three significant challenges of dynamic tasks, heterogeneous resources, and coupled decisions. In this paper, we investigate the issue of joint service caching and computation offloading in mobile edge-cloud computing networks. Specifically, we formulate the optimization problem as minimizing the long-term average service latency, which is NP-hard. To solve the problem, we conduct in-depth theoretical analyses and decompose it into two sub-problems: service caching processing and computation offloading processing. We are the first to propose a novel hierarchical deep reinforcement learning algorithm to solve the formulated problem, where multiple edge agents and a cloud agent collaboratively determine the caching-action and offloading-action, respectively. The results obtained through trace-driven simulations reveal that the proposed framework outperforms several prevailing algorithms concerning the average service latency across diverse scenarios. In a complex real scenario, our framework achieves an approximately 33% convergence improvement and a remarkable 39% reduction in the average service latency when compared to reinforcement learning-based algorithms.