Pruned and Quantized Hybrid Models for Edge-Based Automatic Modulation Recognition

초록

Learning-based automatic modulation recognition on edge devices is envisioned as a critical enabler of real-time spectrum management in future 6G Internet of Things networks. In addition to the requirement of accurate modulation detection, edge devices pose the challenge of reducing the size of the automatic modulation recognition model. This paper proposes a lightweight hybrid model of multi-channel convolutional neural network and the mobile vision transformer. The proposed model addresses both accuracy improvement and model size reduction by employing weight-based pruning and post-training dynamic range quantization. Performance results in edge computing environments such as the Raspberry Pi and Jetson platforms using the RadioML 2016.10a dataset show that the proposed model achieved up to a 91% reduction in memory usage compared to its original version prior to pruning and quantization and demonstrated up to an 8% improvement in average accuracy compared to the baseline convolutional neural network model.

키워드