Transfer Learning for Signal Detection in Wireless Networks

Abstract

The last decade has witnessed the rapid growth of deep learning (DL) applications in wireless communications, especially for channel estimation and signal detection. However, conventional DL techniques are usually trained for a specific scenario. Therefore, it is required to be retrained when applying for new wireless systems. Unfortunately, acquiring enough training data for good detection performance may be costly and time-consuming. Even if the training data is sufficient, conventional DL techniques usually require a long training time. Consequently, this can significantly impact the performance of DL techniques and hinder their robustness. To address all these limitations, this letter introduces a transfer learning framework that can transfer knowledge from a source system to improve the training process of target systems with limited training data. Our simulation results demonstrate that the proposed solution outperforms the conventional deep learning methods in terms of channel estimation and signal detection.