Abstract: This article reviews the main applications, typical learning techniques, and the potential of machine learning (ML) for wireless edge networks. We first analyze the difference between wireless edge intelligence and traditional artificial intelligence. Then, we discuss two approaches to reduce training complexity. One is knowledge-and-data driven ML, which is from the perspective of learning techniques. The other is to design appropriate way of training and decision-making, which is from the perspective of wireless systems. The pros and cons of centralized decision-making and distributed decision-making, centralized training and distributed training are analyzed. We proceed to discuss the application and scope of federated learning for wireless edge networks, and summarize the state-of-the-art and open problems in terms of reducing communication overhead and personalized learning. Finally, we conclude the paper.