Deep learning (DL) has ignited a great research interest in the communications and signal processing society for system design in the physical layer of the wireless communications. DL is particularly advantageous for its model-free nature to provide an efficient way solving the complex and computationally intractable problems. Thus, DL-based works has adopted for the design of next generation communication systems, where intelligent reflecting surfaces (IRSs) are envisioned to be a promising tool due to its hardware efficiency. This article provides an overview of recent advances in the development of DL architectures for IRS-assisted wireless systems. We systematically investigate the DL models for the physical layer design problems, such as signal detection, channel estimation and beamforming. A comprehensive analysis is provided for the advantages and the drawbacks of the DL approaches and related design challenges, such as training data collection, communication- and hardware-efficient model training and environment adaptation, and we highlight the possible future research directions for DL-based wireless communication systems.