Few-Shot Modulation Recognition with Feature Transformation and Meta-Learning
Wendi Xiao, Yuan Zeng, Yi Gong · 2023
Few-shot learning (FSL) has attracted much attention in the fields of image and audio classification, but few efforts have been made on few-shot modulation recognition. It requires classification models to quickly adapt to new modulation recognition tasks with a shift in task distribution. The task distribution shift between training and testing tasks is critical for FSL-based automatic modulation recognition (AMR). In this paper, we propose a few-shot modulation recognition framework using model-agnostic meta-learning (MAML). Our framework includes a single vector in the meta-training structure to mitigate the permutation problem of class label assignments during meta-testing and a feature transformation function in the meta-testing structure to alleviate the channel bias caused by the task distribution shift. We provide an in-depth analysis about the effectiveness of the proposed AMR framework on few-shot modulation recognition. Experimental results demonstrate the proposed framework can greatly improve the generalization ability of learned features and thus the modulation accuracy. Moreover, comparison experiments show that the proposed method outperforms four baseline methods in the few-shot setting.