Auto-encoders in PHY: Principles and Challenges

Deep learning has significantly impacted the physical layer of wireless communications, leading to rapid advancements in this field. Auto-encoder, a self-supervised learning framework, has demonstrated substantial success across various domains. With the continuous improvement of computing power, research on introducing auto-encoders into physical layer wireless communication systems continues to deepen. In this paper, we explain the principles of auto-encoders and explore their applications in communications, particularly in waveform design, joint source-channel coding, and channel representation learning. We further review recent progress, identify key challenges beyond computational limitations, and propose potential future research directions.