An Introduction to Deep Learning for the Physical …

1 1An Introduction to deep Learningfor the Physical LayerTim O Shea,Senior Member, IEEE,and Jakob Hoydis,Member, IEEEA bstract We present and discuss several novel applicationsof deep Learning (DL) for the Physical layer. By interpretinga communications system as an autoencoder, we develop afundamental new way to think about communications systemdesign as an end-to-end reconstruction task that seeks to jointlyoptimize transmitter and receiver components in a single show how this idea can be extended to networks of multipletransmitters and receivers and present the concept of radiotransformer networks (RTNs) as a means to incorporate expertdomain knowledge in the machine Learning (ML) model.

2 Lastly,we demonstrate the application of convolutional neural networks(CNNs) on raw IQ samples for modulation classification whichachieves competitive accuracy with respect to traditional schemesrelying on expert features. The paper is concluded with adiscussion of open challenges and areas for future INTRODUCTIONC ommunications is a field of rich expert knowledge abouthow to model channels of different types [1], [2], compensatefor various hardware imperfections [3], [4], and design optimalsignaling and detection schemes that ensure a reliable transferof data [5].

3 As such, it is a complex and mature engineeringfield with many distinct areas of investigation which haveall seen diminishing returns with regards to performanceimprovements, in particular on the Physical layer. Becauseof this, there is a high bar of performance over which anymachine Learning (ML) or deep Learning (DL) based approachmust pass in order to provide tangible new domains such as computer vision and natural languageprocessing, DL shines because it is difficult to characterize realworld images or language with rigid mathematical models.

4 Forexample, while it is an almost impossible task to write a robustalgorithm for detection of handwritten digits or objects inimages, it is almost trivial today to implement DL algorithmsthat learn to accomplish this task beyond human levels ofaccuracy [6], [7]. In communications, on the other hand,we candesigntransmit signals that enable straightforwardalgorithms for symbol detection for a variety of channel andsystem models ( , detection of a constellation symbol inadditive white Gaussian noise (AWGN)).

5 Thus, as long as suchmodels sufficiently capture real effects we do not expect DLto yield significant improvements on the Physical , we believe that the DL applications whichwe explore in this paper are a useful and insightful way offundamentally rethinking the communications system designT. O Shea is with the Bradley Department of Electrical and ComputerEngineering, Virginia Tech and DeepSig, Arlington, VA, US Hoydis is with Nokia Bell Labs, Route de Villejust, 91620 Nozay, and hold promise for performance improvementsin complex communications scenarios that are difficult todescribe with tractable mathematical models.

6 Our main con-tributions are as follows: We demonstrate that it is possible to learn full transmitterand receiver implementations for a given channel modelwhich are optimized for a chosen loss function ( ,minimizing block error rate (BLER)). Interestingly, such learned systems can be competitive with respect to thecurrent state-of-the-art. The key idea here is to representtransmitter, channel, and receiver as one deep neuralnetwork (NN) that can be trained as an autoencoder. Thebeauty of this approach is that it can even be applied tochannel models and loss functions for which the optimalsolutions are unknown.

7 We extend this concept to an adversarial network ofmultiple transmitter-receiver pairs competing for leads to the interference channel for which findingthe best signaling scheme is a long-standing researchproblem. We demonstrate that such a setup can also berepresented as an NN with multiple inputs and outputs,and that all transmitter and receiver implementationscan be jointly optimized with respect to a common orindividual performance metric(s). We introduce radio transformer networks (RTNs) as away to integrate expert knowledge into the DL allow, for example, to carry out predefined cor-rection algorithms ( transformers ) at the receiver ( ,multiplication by a complex-valued number, convolutionwith a vector) which may be fed with parameters learnedby another NN.

8 This NN can be integrated into theend-to-end training process of a task performed on thetransformed signal ( , symbol detection). We study the use of NNs on complex-valued IQ samplesfor the problem of modulation classification and showthat convolutional neural networks (CNNs), which arethe cornerstone of most DL systems for computer vision,can outperform traditional classification techniques basedon expert features. This result mirrors a relentless trendin DL for various domains, where learned features ulti-mately outperform and displace long-used expert features,such as the scale-invariant feature transform (SIFT) [8]and Bag-of-words [9].

9 The ideas presented in this paper provide a multitude ofinteresting avenues for future research that will be discussedin detail. We hope that these will stimulate wide interest withinthe research [ ] 11 Jul 20172 The rest of this article is structured as follows: Section I-Adiscusses potential benefits of DL for the Physical layer. Sec-tion I-B presents related work. Background of deep Learning ispresented in Section II. In Section III, several DL applicationsfor communications are presented. Section IV contains anoverview and discussion of open problems and key areas offuture investigation.

10 Section V concludes the Potential of DL for the Physical layerApart from the intellectual beauty of a fully learned communications system, there are some reasons why DL couldprovide gains over existing Physical layer , most signal processing algorithms in communicationshave solid foundations in statistics and information theory andare often provably optimal for tractable mathematically mod-els. These are generally linear, stationary, and have Gaussianstatistics. A practical system, however, has many imperfectionsand non-linearities [4] ( , non-linear power amplifiers (PAs),finite resolution quantization) that can only be approximatelycaptured by such models.