Influence of Autoencoder-Based Data Augmentation on Deep Learning-Based Wireless Communication

Abstract

Deep learning (DL) has been gradually applied to wireless communication and has achieved remarkable results. However, training a DL model requires numerous data, and an insufficient training dataset will cause serious overfitting problem and reduce model accuracy. Data augmentation (DA) is one of the commonly used techniques to solve the above problem. It is widely used to improve the performance of image and text classification tasks, but its impact on DL-based wireless communication has not been fully explored. Inspired by the emerging deep autoencoder (AE) generative model, we propose an AE-based DA method to improve the generalization performance of the DL-based wireless communication. We perform experimental verification on the two tasks of channel estimation model of wireless communication physical layer and massive multiple-input multiple-output (MIMO) power allocation optimization model. Experimental results show that the AE-based DA method can improve the generalization performance of the wireless communication, but this improvement is related to the training dataset size. When the training dataset size is smaller than a threshold, AE can improve model performance by increasing the relevant training data, but when it is larger than this threshold, it reduces model performance. We also propose that the straight line intersection method can be used directly to roughly determine this threshold. Furthermore, we propose a mixup-based method to solve the problem that the AE cannot improve model performance when the training dataset size is larger than the threshold.