Abstract
This paper proposes a resource allocation strategy for dynamic training and inference of machine learning tasks at the edge of the wireless network. The goal is to explore and optimize the fundamental trade-offs between energy consumption, end-to-end (E2E) service delay, and learning accuracy. Utilizing stochastic Lyapunov optimization, the authors develop low-complexity algorithms that manage radio and computation resources in real-time without requiring prior knowledge of network statistics.
The Challenge of Wireless Edge Machine Learning
As the industry moves toward 6G and advanced IoT, there is a significant shift from centralized cloud computing to Edge Machine Learning (EML). In this paradigm, mobile devices stream data to nearby edge servers to perform training or inference tasks. However, the wireless edge is a resource-constrained environment where mobile devices have limited battery life, wireless channels are volatile, and edge servers have finite computational capacity.
Unlike traditional computation offloading that treats tasks as binary decisions, this research treats learning accuracy as a controllable variable. By adjusting data quantization levels before transmission, the system can save energy and time at the cost of model precision. The core challenge lies in dynamically finding the "sweet spot" among these three competing metrics.
A Dynamic Resource Allocation Framework
The proposed framework is based on Stochastic Lyapunov Optimization, a technique that allows a system to satisfy long-term average constraints by making optimal decisions in short, discrete time slots.
The architecture manages several "Resource Knobs":
- Transmission Data Rate: Controls how fast bits are sent over the air.
- Quantization Levels: Determines how many bits represent each data sample, serving as the primary lever for accuracy.
- CPU Scheduling: Allocates edge server clock frequency across different users.
The algorithm maintains virtual queues to track deviations from energy, delay, and accuracy targets. If a queue grows too long, the system prioritizes that constraint in the next time slot. This data-driven approach requires no prior knowledge of future data arrivals or channel conditions.
Key Insights: Quantization as a Control Lever
A significant contribution of this work is treating the number of quantization bits not merely as a compression tool, but as a direct proxy for learning accuracy. This allows the network to "downgrade" accuracy slightly to prevent massive spikes in latency or energy during poor channel conditions. This "goal-oriented" communication prioritizes the final decision or estimate over perfect bit transmission.
The authors introduce a control parameter ($V$) that acts as a dial. Increasing $V$ forces the system to prioritize energy savings, which naturally leads to higher latency and lower accuracy.
Experimental Results
The framework was validated using synthetic regression data and real-world datasets, including a Gas Sensor Array, MNIST handwritten digits, and Hydraulic System Monitoring.
- Trade-off Curves: Results demonstrate that relaxing accuracy requirements (e.g., allowing higher Mean Squared Error) significantly reduces E2E delay and sensor energy consumption.
- Convergence: In gas sensor experiments, the algorithm successfully converged to target accuracy levels while maintaining a steady 20ms average delay.
- Classification Performance: For SVM and Neural Network tasks, the framework maximized the "Correct Classification Rate" while strictly adhering to device energy budgets. Power-constrained devices automatically scaled back quantization to preserve battery life, while those with ample power achieved higher accuracy.
Conclusion
This research provides a robust, low-complexity solution for managing the complex trade-offs inherent in Edge Machine Learning. By integrating learning accuracy into the resource allocation problem, the framework enables more efficient and goal-oriented communication in future wireless networks. Future work may explore more sophisticated vector quantizers and non-linear CPU models for complex deep learning architectures.
Summary updated