Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 92 of the 92 references it cites were found in a registry or an index; 8 of the 32 sources its claims cite were checked for support.

An Introduction to Deep Learning for the Physical Layer

Open access
O'Shea, Timothy, Hoydis, Jakob · 2017 · IEEE Transactions on Cognitive Communications and Networking

Abstract

We present and discuss several novel applications of deep learning for the physical layer. By interpreting a communications system as an autoencoder, we develop a fundamental new way to think about communications system design as an end-to-end reconstruction task that seeks to jointly optimize transmitter and receiver components in a single process. We show how this idea can be extended to networks of multiple transmitters and receivers and present the concept of radio transformer networks as a means to incorporate expert domain knowledge in the machine learning model. Lastly, we demonstrate the application of convolutional neural networks on raw IQ samples for modulation classification which achieves competitive accuracy with respect to traditional schemes relying on expert features. This paper is concluded with a discussion of open challenges and areas for future investigation.

Overview

CoChat Summary

TL;DR

The paper introduces deep learning for the physical layer by treating a communications system as an autoencoder that jointly learns transmitter and receiver implementations end-to-end. It also extends this idea to multi-user interference channels, proposes radio transformer networks (RTNs) to inject expert domain knowledge, and shows CNNs on raw IQ samples can outperform expert-feature modulation classifiers.

Key findings

  • A learned autoencoder for a (7,4) system at rate R = 4/7 achieves the same block error rate (BLER) as a Hamming (7,4) code with maximum likelihood decoding, while an (8,8) autoencoder outperforms uncoded BPSK over the full E_b/N_0 range, indicating learned joint coding and modulation.
  • For the two-user AWGN interference channel, the autoencoder matches time-sharing for (1,1) and (2,2), but achieves gains of about 0.7 dB for (4,4) and 1 dB for (4,8) at a BLER of 10^-3.
  • An autoencoder (8,4) with an RTN over a multipath fading channel with L = 3 Rayleigh taps outperforms differential BPSK with maximum likelihood sequence estimation and a Hamming (7,4) code, whereas the plain autoencoder struggles to match it; RTNs also show faster training convergence.
  • For modulation classification on 1.2M sequences of 128 complex-valued IQ samples across ten modulation schemes and SNRs from -20 dB to 18 dB, the CNN outperforms a boosted expert-feature classifier by around 4 dB in the low-to-medium SNR range, with almost identical high-SNR performance; a single tree is about 6 dB worse at medium SNR and 3.5% worse at high SNR.

Methods

  • The transmitter, channel, and receiver are represented as one deep neural network trained as an autoencoder: one-hot message input, dense layers, normalization layer for transmit constraints, additive noise channel layer, and softmax receiver output, trained with categorical cross-entropy using SGD/Adam at learning rate 0.001 and fixed E_b/N_0 = 7 dB.
  • The autoencoder concept is extended to multiple transmitter-receiver pairs in a two-user interference channel, with dynamic loss weights α_t per mini-batch to balance the two cross-entropy losses; performance is compared against 2^{2k/n}-QAM with time-sharing.
  • RTNs combine a learned parameter estimator, a differentiable parametric transform, and a discriminative network; the paper demonstrates phase-offset compensation and multipath inverse filtering. For modulation classification, a VGG-like CNN with 324,330 trainable parameters is trained on the RML2016.10b dataset.

Limitations

  • The paper states that scalability to large message sets M is a major challenge: for k = 100 bits, training complexity is prohibitive because the autoencoder must see at least every message once.
  • The paper states that optimal data representations, loss functions, training SNR, and hyperparameters are not well understood; training at one SNR does not necessarily generalize to others, and no common communications benchmarks or datasets comparable to MNIST or ImageNet exist.
  • The paper states that end-to-end learning over real channels requires system identification because backpropagation assumes a known channel transfer function; complex-valued neural networks are not supported by current DL libraries. Evident from the text: the reported results are simulation-based, use small (n,k) settings, assume known channel models during training, and the paper explicitly leaves detailed performance comparisons for various channel types and parameters to future work.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tccn.2017.2758370

Cites

92 of 92 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 2 A communications system over an AWGN channel represented as an autoencoder. The input s is encoded as a one-hot vector, the output is a probability distribution over all possible messages from which the most likely is picked as output ˆ s .
Figure 2 A communications system over an AWGN channel represented as an autoencoder. The input s is encoded as a one-hot vector, the output is a probability distribution over all possible messages from which the most likely is picked as output ˆ s .From the PDF
Figure 3 BLER versus E b /N 0 for the autoencoder and several baseline communication schemes
Figure 3 BLER versus E b /N 0 for the autoencoder and several baseline communication schemesFrom the PDF
Figure 4 Constellations produced by autoencoders using parameters ( n, k ) : (a) (2 , 2) (b) (2 , 4) , (c) (2 , 4) with average power constraint, (d) (7 , 4) 2-dimensional t-SNE embedding of received symbols.
Figure 4 Constellations produced by autoencoders using parameters ( n, k ) : (a) (2 , 2) (b) (2 , 4) , (c) (2 , 4) with average power constraint, (d) (7 , 4) 2-dimensional t-SNE embedding of received symbols.From the PDF
Figure 5 The two-user interference channel seen as a combination of two interfering autoencoders that try to reconstruct their respective messages
Figure 5 The two-user interference channel seen as a combination of two interfering autoencoders that try to reconstruct their respective messagesFrom the PDF
Figure 6 BLER versus E b /N 0 for the two-user interference channel achieved by the autoencoder (AE) and 2 2 k/n -QAM time-sharing (TS) for different parameters ( n, k )
Figure 6 BLER versus E b /N 0 for the two-user interference channel achieved by the autoencoder (AE) and 2 2 k/n -QAM time-sharing (TS) for different parameters ( n, k )From the PDF
Figure 7 Learned constellations for the two-user interference channel with parameters (a) (1 , 1) , (b) (2 , 2) , (c) (4 , 4) , and (d) (4 , 8) . The constellation points of Transmitter 1 and 2 are represented by red dots and black crosses, respectively.
Figure 7 Learned constellations for the two-user interference channel with parameters (a) (1 , 1) , (b) (2 , 2) , (c) (4 , 4) , and (d) (4 , 8) . The constellation points of Transmitter 1 and 2 are represented by red dots and black crosses, respectively.From the PDF
Figure 8 A radio receiver represented as an RTN. The input y fi rst runs through a parameter estimation network g ω ( y ) , has a known transform t ( y , ω ) applied to generate the canonicalized signal y , and then is classified in the discriminative network g ( y ) to produce the output ˆ s .
Figure 8 A radio receiver represented as an RTN. The input y fi rst runs through a parameter estimation network g ω ( y ) , has a known transform t ( y , ω ) applied to generate the canonicalized signal y , and then is classified in the discriminative network g ( y ) to produce the output ˆ s .From the PDF
Figure 9 BLER versus E b /N 0 for various communication schemes over a channel with L = 3 Rayleigh fading taps
Figure 9 BLER versus E b /N 0 for various communication schemes over a channel with L = 3 Rayleigh fading tapsFrom the PDF
Figure 10 Autoencoder training loss with and without RTN
Figure 10 Autoencoder training loss with and without RTNFrom the PDF
Figure 11 Classifier performance comparison versus SNR
Figure 11 Classifier performance comparison versus SNRFrom the PDF
Figure 12 Confusion matrix of the CNN (SNR = 10 dB)
Figure 12 Confusion matrix of the CNN (SNR = 10 dB)From the PDF

Verification

Claims the paper makes

6 Valid · 4 Not valid · 18 Not checked · 65 Own results and interpretations

39 more claims

Create a free account to see all 93 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 29
  • It may state two separate claims as one. 24

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 69 references

Research works printed but not linked: 5

  1. goodfellow, 2016, Deep Learning
  2. estaran, 2016, Artificial neural networks for linear and non-linear impairment mitigation in high-baudrate IM/DD systems, Proc 42nd Eur Conf Opt Commun (ECOC), 1
  3. al-rfou, 2016, Theano: A Python framework for fast computation of mathematical expressions, arXiv preprint arXiv 1605 01584
  4. o’shea, 2017, Source Code
  5. chollet, 2015, Keras

Cited by

See all 2,994

Metadata

Authors
O'Shea, T., & Hoydis, J.
Affiliations
Nokia (France) (FR)
Venue
IEEE Transactions on Cognitive Communications and Networking
Published
2017-12
Volume
3
Issue
4
Pages
563-575
ISSN
2332-7731 (ISSN-L)
OpenAlex
W2734408173
Cited by
2,980
References
69
Citations by year
2026: 307 · 2025: 327 · 2024: 305 · 2023: 368 · 2022: 432 · 2021: 409 · 2020: 406 · 2019: 322 · 2018: 95 · 2017: 9
FWCI
122.29
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Cognitive Communications and Networking (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Wireless Signal Modulation Classification
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Wireless Communication Techniques
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Wireless Communication Technologies
Keywords
end-to-end reconstruction, domain knowledge integration, convolutional neural networks, automatic modulation classification, IQ samples

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-11-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: An Introduction to Deep Learning for the Physical Layer. https://app.cochat.ai/papers/10.1109/tccn.2017.2758370
The paper
O'Shea, T., & Hoydis, J. (2017). An Introduction to Deep Learning for the Physical Layer. IEEE Transactions on Cognitive Communications and Networking. https://doi.org/10.1109/tccn.2017.2758370

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.