Power Control for NN-based Wireless Distributed Inference with Improved Model Calibration
Qiushuo Hou,Mengyuan Lee,Guanding Yu,Yunlong Cai
DOI: https://doi.org/10.1109/twc.2024.3417305
2024-01-01
Abstract:In recent years, the application of neural networks (NNs) in wireless communication has garnered widespread attention and proven successful. However, conventional learning-based NNs often suffer from poor calibration, meaning that they struggle to reliably quantify prediction confidence and lack proper uncertainty estimation. This limitation becomes especially critical for next generation communication systems, particularly in complex industrial scenarios with stringent reliability requirements. Previous efforts to enhance model calibration have primarily centered on modifying NNs’ training processes. However, these methods often demand significant computing resources, making them impractical for resource-constrained scenarios. In this paper, we investigate a distributed wireless communication system involving multiple users and propose a novel approach to improve model calibration. Our method focuses on enhancing calibration during the inference stage of NNs by introducing a power control mechanism. Notably, existing research indicates that many NNs exhibit overconfidence, i.e., the NN’s confidence exceeds its actual accuracy. Leveraging this insight, we exploit the inherent noise and fading in wireless systems to naturally reduce the NN’s confidence while preserving accuracy. To achieve this, we employ linear relaxation-based perturbation analysis (LiRPA) to approximate the relationship between the perturbed output and the input perturbation of the NN. Subsequently, we devise an optimization problem by leveraging the analyzed relationship and the definition of perfect calibration. It is aimed at finding the input perturbation that maximizes the probability of the model achieving perfect calibration. Finally, considering different channel conditions and a given specific modulation method, we derive the optimal transmission power based on bit error rate (BER) formula. Simulation results demonstrate that our proposed power control method exhibits significant advantages in model calibration compared to several traditional approaches.