This paper introduces novel applications of deep learning (DL) for the physical layer of communications. By interpreting a communications system as an autoencoder, the authors propose a new approach to system design as an end-to-end reconstruction task that jointly optimizes transmitter and receiver components. They extend this idea to networks of multiple transmitters and receivers, introducing the concept of radio transformer networks (RTNs) to incorporate expert domain knowledge into the machine learning (ML) model. The paper also demonstrates the application of convolutional neural networks (CNNs) on raw IQ samples for modulation classification, achieving competitive accuracy with traditional schemes. The authors discuss the potential of DL for the physical layer, highlighting that while traditional methods are based on solid statistical and information-theoretic foundations, DL can potentially better handle practical imperfections and non-linearities in real-world systems. They argue that DL can provide significant improvements in complex communication scenarios that are difficult to model with traditional methods. The paper also discusses open challenges and areas for future research, including the need for common benchmarks and datasets, optimal data representations, loss functions, and training strategies for DL in communications. The paper presents several DL applications for communications, including the use of autoencoders for end-to-end communications systems, the extension of this concept to multiple transmitters and receivers, and the application of RTNs for augmented signal processing algorithms. The authors also demonstrate the effectiveness of CNNs for modulation classification, showing that they can outperform traditional classification techniques based on expert features. The paper concludes with a discussion of open research challenges and future directions in the application of DL to the physical layer of communications.This paper introduces novel applications of deep learning (DL) for the physical layer of communications. By interpreting a communications system as an autoencoder, the authors propose a new approach to system design as an end-to-end reconstruction task that jointly optimizes transmitter and receiver components. They extend this idea to networks of multiple transmitters and receivers, introducing the concept of radio transformer networks (RTNs) to incorporate expert domain knowledge into the machine learning (ML) model. The paper also demonstrates the application of convolutional neural networks (CNNs) on raw IQ samples for modulation classification, achieving competitive accuracy with traditional schemes. The authors discuss the potential of DL for the physical layer, highlighting that while traditional methods are based on solid statistical and information-theoretic foundations, DL can potentially better handle practical imperfections and non-linearities in real-world systems. They argue that DL can provide significant improvements in complex communication scenarios that are difficult to model with traditional methods. The paper also discusses open challenges and areas for future research, including the need for common benchmarks and datasets, optimal data representations, loss functions, and training strategies for DL in communications. The paper presents several DL applications for communications, including the use of autoencoders for end-to-end communications systems, the extension of this concept to multiple transmitters and receivers, and the application of RTNs for augmented signal processing algorithms. The authors also demonstrate the effectiveness of CNNs for modulation classification, showing that they can outperform traditional classification techniques based on expert features. The paper concludes with a discussion of open research challenges and future directions in the application of DL to the physical layer of communications.