Abstract: In recent years, machine learning technology represented by deep learning has been developing rapidly. With its excellent learning ability, it has shown its unique advantages in modeling problems under complex environmental conditions. Currently, the research on the machine learning based underwater acoustic (UWA) communication is flourishing, and certain progresses have been made in channel estimation and equalization, typical communication system applications, and other aspects. However, few researches regard the real-world constraints on UWA communications. Therefore, the development ideas of UWA channel estimation with data augmentation, label- free learning, and few-shot learning are discussed in view of the issues of insufficient samples in UWA communication, labeling difficulties, and source/target domain mismatches due to the spatial and temporal variability. Besides, the preliminary simulation and experimental results are given. This paper is a preliminary exploration of the important and difficult issues in the intersection research on UWA communication and machine learning, which provides a reference for the development of autonomous and intelligent communication technology of various underwater platforms.