Abstract
As AI applications migrate to the network edge, traditional Mobile Edge Computing (MEC) offloading faces significant hurdles due to massive data volumes and the non-linear computational demands of deep learning models. This paper proposes a novel framework that jointly optimizes semantic transmission and early exit of inference (EEoI). By extracting semantic features to reduce uplink traffic and utilizing multi-exit neural networks to conserve computational cycles, the system significantly improves resource utilization and Quality of Experience (QoE). Experimental results demonstrate that the proposed joint optimization outperforms traditional linear-model-based offloading in both latency and system reward.
1. Introduction
The proliferation of AI-driven edge applications, such as autonomous driving and real-time image recognition, has strained existing MEC infrastructures. Traditional offloading methods struggle with two primary issues: the massive uplink bandwidth required for high-resolution data and the complex, non-linear computational costs of Deep Neural Networks (DNNs). Unlike traditional tasks, DNN complexity does not scale linearly with data size, making simple bit-to-cycle ratios inadequate for resource management. This paper addresses these challenges by integrating intelligent communication and computation mechanisms.
2. System Architecture and Methodology
The proposed framework leverages two key "intelligent" mechanisms to streamline the offloading process:
- Semantic Transmission: Instead of transmitting raw pixel data, terminals utilize a convolutional neural network (CNN) to extract and compress essential "semantic" features. This process, illustrated in the semantic extraction and compression blocks, drastically reduces the data volume sent over the wireless uplink.
- Early Exit of Inference (EEoI): AI models deployed on MEC servers are designed with multiple "exit points." If a task can be classified with high confidence at an early layer, the inference process terminates early, saving significant computational resources.
3. Joint Optimization Framework
To manage the interplay between communication and computation, the authors formulate a resource allocation problem aimed at maximizing a reward-based QoE. The problem is solved using a Block Coordinate Descent (BCD) framework, which decomposes the optimization into three subproblems:
- Communication Resource Allocation: A 3D matching algorithm associates terminals with base stations and subcarriers.
- Semantic Transmission Selection: The system dynamically decides between semantic and direct transmission for each task.
- Computation Capacity Allocation: Convex optimization is used to distribute MEC server power among tasks.
A critical innovation is the use of FLOP (Floating Point Operation) counts to model computational complexity, providing a hardware-accurate representation of DNN performance compared to traditional linear models. The variables and parameters governing this framework are detailed in the notation table below.
4. Performance Evaluation
The framework was validated using the CIFAR-10 dataset, employing DeepJSCC-V for semantic transmission and a modified B-Alexnet for the EEoI mechanism. Simulations in a multi-base station environment yielded the following results:
- Latency Reduction: The proposed algorithm achieved a 4.3% to 4.8% improvement in latency over models relying on traditional linear computation estimates.
- System Reward: The joint method increased the weighted system reward by 22.5% compared to fixed transmission modes and by 6.6% compared to nearest-base-station association strategies.
- EEoI Impact: Increasing the number of exit points from 1 to 3 improved the average terminal reward by approximately 10.9%, highlighting the efficiency of early inference termination.
5. Conclusion and Future Work
This research demonstrates a powerful synergy between semantic communication and adaptive computation. By treating communication and computation as an integrated pipeline, the system achieves higher efficiency than siloed optimizations. Future research will focus on supporting heterogeneous AI model architectures, exploring the Pareto front between inference accuracy and delay, and adapting the framework for highly dynamic, high-mobility environments.
Summary updated