Machine learning (ML)-based communication systems are a promising technology for 5G and beyond wireless communication systems. ML-based approaches can discover inherent linear or nonlinear characteristics from sufficient amount of data, which can be applied to wireless communication systems. As the structure of wireless communication systems is becoming more complex, designing optimal channel estimators and symbol detectors is extremely challenging, often impossible. Surprisingly, it has been shown that a deep neural network (DNN), e.g., deep convolutional neural network (CNN) or multi-layer perceptron (MLP), can achieve nearly optimal channel estimation and symbol detection performance. Also, wireless communications-based ML framework introduces various interesting systems that differ from the conventional systems, such as over-the-air federated learning systems. To make ML-based communication systems practical, however, the large training overhead and overfitting must be resolved, which require extensive research efforts.