Introduction
Modern Deep Learning (DL) models are computationally intensive, posing significant challenges for deployment on resource-constrained IoT devices. While offloading tasks to edge servers can mitigate local constraints, fluctuating wireless channel conditions and latency requirements necessitate a more nuanced approach. This paper explores "goal-oriented" or semantic communication, where the primary objective is successful task completion (e.g., accurate classification) rather than perfect bit reconstruction. We propose a framework that dynamically balances local computation and edge offloading using recursive early exit neural networks and Reinforcement Learning (RL).
System Architecture and Methods
The proposed system consists of two main components: a Recursive Early Exit (EE) architecture and an RL-based optimization agent.
Recursive Early Exit (EE) Architecture
Unlike traditional early-exit models with independent classifiers, our recursive model integrates features from the current layer with predictions from previous exits. This approach utilizes "positive and negative moving mass" functions to refine class probabilities progressively. This recursive refinement ensures that confidence builds logically across layers, improving parameter efficiency and prediction stability.
RL-Based Online Optimization
A Q-learning agent manages the decision-making process at each inference step. Based on the current exit point and wireless channel quality (MCS), the agent selects one of three actions:
- Exit: Terminate and use the current local prediction.
- Compute: Proceed to the next local neural network layer.
- Offload: Transmit intermediate features to the edge server for final processing.
The agent is rewarded based on a "Goal Effectiveness" metric, which requires both classification accuracy (determined by a confidence margin) and completion within a strict deadline ($D_{max}$).
Experimental Setup
The framework was evaluated using the CIFAR-10 dataset with a ResNet-20 backbone modified to include nine recursive early exit points. The wireless environment was simulated within a 100m radius, accounting for Rayleigh fading and path loss to model realistic channel fluctuations.
Results and Discussion
The recursive model demonstrates a superior trade-off between inference accuracy and computational cost (FLOPs) compared to standard halting strategies like "highest probability" or "patience-based" halting.
Key findings include:
- Efficiency: The proposed model achieves higher accuracy for the same computational budget.
- Adaptability: The RL agent effectively learns to compute locally when channel conditions are poor and offload early when high-speed transmission is available, ensuring latency constraints are met.
- Goal Effectiveness: The system maintains high accuracy (~85% on CIFAR-10) while strictly adhering to a 40ms deadline.
Conclusion
This research highlights the potential of combining recursive neural architectures with RL for goal-oriented communication in 6G networks. By prioritizing task success and resource efficiency, the framework enables robust DL performance on IoT devices. Future work will explore multi-agent scenarios and more complex tasks beyond image classification.
Summary updated