Large Models for Aerial Edges: An Edge-Cloud Model Evolution and Communication Paradigm

Shuhang Zhang, Qingyu Liu, Ke Chen, Boya Di, Hongliang Zhang, Wenhan Yang, Dusit Niyato, Zhu Han, H. Vincent Poor

TL;DR

This paper proposes a novel framework for integrating large AI models on aerial edge devices (like drones) with cloud computing resources. It tackles the challenge of limited aerial computing power and communication bandwidth by developing a joint optimization strategy for task allocation, data transmission, and model updates, significantly improving inference accuracy.

What Problem Does This Paper Address?

The advent of 6G wireless networks promises ubiquitous connectivity and advanced applications powered by large AI models. However, deploying these massive models on resource-constrained aerial platforms like drones (UAVs) presents a significant hurdle. UAVs lack the computational power and storage for training or even running large models, leading to degraded inference accuracy in dynamic environments. Existing approaches often offload all computation to the cloud, which can be inefficient due to communication bottlenecks, or rely solely on edge devices, which limits model complexity. This paper addresses this gap by proposing an integrated air-ground edge-cloud framework that intelligently balances computation between UAVs and the cloud, enabling model evolution and improved performance.

Methods

The core of this work is an integrated air-ground edge-cloud model evolution framework. This framework allows UAVs to act as edge nodes for data collection and initial processing using smaller, onboard models. They can then selectively upload extracted features and "residual mapping data" (supplementary visual information) to a ground cloud server for analysis with larger, more powerful models. Crucially, the cloud server also provides model updates back to the UAVs, enabling their onboard models to evolve and improve over time.

The authors formulate a joint optimization problem to maximize the overall Mean Average Precision (mAP) of this integrated system. This involves simultaneously optimizing:

  • Task Allocation: Deciding which frames are processed by the edge UAV and which are sent to the cloud.
  • Transmission Resource Allocation: Dynamically allocating uplink and downlink bandwidth for feature streams, data streams (residual mapping), and model streams.
  • Data Quantization: Determining the optimal bit precision for residual mapping data to balance accuracy and bandwidth usage.
  • Edge Model Update Design: Managing the overhead of model updates transmitted from the cloud to the UAV.

To solve this complex problem, they decompose it into two subproblems: a data stream design subproblem (optimizing residual mapping data transmission and quantization) and a feature/model stream design subproblem (optimizing task allocation, bandwidth, and model updates). They also derive theoretical bounds and approximations for the mAP to facilitate optimization.

Datasets

The paper uses vision-based classification experiments for evaluation. Specifically, the simulations are based on results from a classification task on the CIFAR10 dataset. The models used are a ResNet18 (11.7M parameters) for the edge UAV and a ResNet101 (45M parameters) for the cloud server. For further analysis, they also consider larger models like ResNet34 (20M parameters) for the UAV and ResNet101 (110M parameters) for the cloud.

Key Insights and Ideas

  • Intelligent Data Transmission is Key: Not all data needs to be sent to the cloud. Extracting features at the edge and selectively sending "residual mapping data" with adjustable quantization is crucial for efficient communication.
  • Model Evolution is Essential: The framework isn't static; it allows edge models to learn and improve from cloud model updates, enhancing their performance in dynamic environments.
  • Dynamic Resource Allocation Adapts to Bandwidth: The system intelligently shifts computation and data transmission strategies based on available wireless bandwidth, prioritizing edge processing when bandwidth is low and cloud processing when it's high.
  • Trade-offs Exist: There's a fundamental trade-off between edge model performance (mAPs) and cloud model performance (mAPL), and the framework aims to find the optimal balance. Uplink spectrum efficiency has a more significant impact on mAP than downlink efficiency due to the larger overhead of residual mapping data.

Key Results

The proposed integrated air-ground edge-cloud model evolution framework consistently outperforms a centralized cloud-only model and a distributed edge-only model across various communication bandwidths and data sizes. The performance gap to an exhaustive search (considered an upper bound) is minimal (0-0.5%).

  • Bandwidth Impact: As total bandwidth increases, the framework increasingly relies on the cloud for higher mAP analysis. With very low bandwidth (< 2 MHz), the edge model's contribution becomes more significant.
  • Frame Rate Impact: Increasing the number of frames per second (N) leads to a decrease in mAP, as less residual mapping data can be transmitted per frame. However, the framework's mAP remains bounded by the edge model's performance.
  • Spectrum Efficiency: Both uplink and downlink spectrum efficiencies positively correlate with mAP. Uplink efficiency is more critical for transmitting residual mapping data.

Limitations and Future Work

  • Assumptions: The analysis relies on assumptions about the concavity of mAP functions with respect to quantization bits and model updates, which are based on experimental observations rather than strict theoretical proofs.
  • Single UAV Scenario: The current system model focuses on a single UAV. Future work will explore resource allocation among multiple UAVs in a multi-UAV scenario.
  • Spectrum Efficiency Optimization: The paper treats spectrum efficiencies (Su and Sd) as constants. Future research will investigate optimizing these factors, considering aspects like UAV trajectory, beamforming, and interference management.
  • Model Update Overhead: While the framework considers model update overhead, further research could explore more advanced model compression and incremental updating techniques.

Summary updated

Abstract

The future sixth-generation (6G) of wireless networks is expected to surpass its predecessors by offering ubiquitous coverage through integrated air-ground facility deployments in both communication and computing domains. In this network, aerial facilities, such as unmanned aerial vehicles (UAVs), conduct artificial intelligence (AI) computations based on multi-modal data to support diverse applications including surveillance and environment construction. However, these multi-domain inference and content generation tasks require large AI models, demanding powerful computing capabilities, thus posing significant challenges for UAVs. To tackle this problem, we propose an integrated edge-cloud model evolution framework, where UAVs serve as edge nodes for data collection and edge model computation. Through wireless channels, UAVs collaborate with ground cloud servers, providing cloud model computation and model updating for edge UAVs. With limited wireless communication bandwidth, the proposed framework faces the challenge of information exchange scheduling between the edge UAVs and the cloud server. To tackle this, we present joint task allocation, transmission resource allocation, transmission data quantization design, and edge model update design to enhance the inference accuracy of the integrated air-ground edge-cloud model evolution framework by mean average precision (mAP) maximization. A closed-form lower bound on the mAP of the proposed framework is derived, and the solution to the mAP maximization problem is optimized accordingly. Simulations, based on results from vision-based classification experiments, consistently demonstrate that the mAP of the proposed framework outperforms both a centralized cloud model framework and a distributed edge model framework across various communication bandwidths and data sizes.
BibTeX
@article{zhang2024large,
  author = {Shuhang Zhang and Qingyu Liu and Ke Chen and Boya Di and Hongliang Zhang and Wenhan Yang and Dusit Niyato and Zhu Han and H. Vincent Poor},
  title = {{Large Models for Aerial Edges: An Edge-Cloud Model Evolution and Communication Paradigm}},
  journal = {IEEE J. Sel. Areas Commun.},
  year = {2024},
  doi = {10.1109/jsac.2024.3460078},
}
Figure