Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Bandwidth-Agile Image Transmission with Deep Joint Source-Channel Coding

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A single learned encoder can split an image into wireless layers that reconstruct almost as well as one direct transmission, for both ordered and unordered layer reception.

desk verdict A clearly written, honest empirical extension of DeepJSCC to layered and multiple-description transmission; the main result is plausible and useful, but 'negligible loss' is not yet quantified and the high-resolution/many-layer case is untested. read the letter →

arxiv 2009.12480 v2 pith:RRE5NHGF submitted 2020-09-26 cs.IT cs.LGeess.IVeess.SPmath.IT

David Burth Kurka , Deniz Gündüz This is my paper · ORCID
classification cs.ITcs.LGeess.IVeess.SPmath.IT
keywords jointsource-channelcodingsuccessiverefinementmultipledescriptionswirelessimagetransmissiondeeplearningbandwidth-agilecommunicationgracefuldegradationconvolutionalautoencoder
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes DeepJSCC-l, a way to transmit an image over several independent wireless channels so that a receiver can use any subset of the received channel outputs and still reconstruct the image. The transmitter runs one convolutional encoder that emits all layers at once; each possible subset of layers has its own decoder, and the whole system is trained end to end with the noisy channel inserted as a fixed differentiable layer. Experiments on CIFAR-10 show that decoding only the first layers gives a useful image, while adding layers improves quality by roughly the same amount as a single transmission that used the same total bandwidth from the start. The same design covers both ordered layers (successive refinement) and unordered independent layers (multiple descriptions). If the result holds, wireless image delivery can adapt its bandwidth on the fly, letting receivers tune into more or fewer layers depending on their channel or quality target without a separate compression-and-coding pipeline.

What carries the argument

The machine that carries the result is DeepJSCC-l's autoencoder: a single CNN encoder f_theta maps the image to one vector z=(z1,...,zL) that is split across L channels, and one decoder g_{phi_S} per receivable subset S reconstructs from the noisy outputs; every decoder shares the same encoder parameters theta, and the channel is a non-trainable differentiable layer so gradients flow through noise. Training minimizes the average MSE over all subset reconstructions, which forces the encoder to arrange information so that every prefix (successive refinement) or every combination (multiple descriptions) is useful. Alternative architectures trade memory and training time: a single decoder trained with random zero-masking of missing layers, and a residual scheme where later encoder-decoder pairs are trained greedily on an estimate of the receiver's reconstruction error.

What would settle it

Train DeepJSCC-l with L=2 and total bandwidth ratio 1/6 on a natural-image dataset, then compare its two-layer reconstruction PSNR to a single-transmission DeepJSCC model of the same total bandwidth ratio at the same SNR. If the layered curve falls below the single-transmission curve by more than the small margin shown for CIFAR-10, and the gap grows when L is raised to 8 or 16, the claim that layering is nearly lossless fails outside the tested setting.

Watch

Extended reading notes

Core claim

DeepJSCC-l is the claim that one jointly trained convolutional encoder, shared by all layers, can learn a layered representation of a natural image that is nearly successively refinable over Gaussian channels: the reconstruction from the first i layers costs almost none of the end-to-end quality that a dedicated single-layer encoder of the same total bandwidth would give. For L=2, 5, and up to 16 layers at total bandwidth ratio 1/3, the layered PSNR curves stay close to the single-transmission baselines, and the first layers' quality barely changes as more refinement layers are added. The paper also demonstrates the harder multiple-description version, where any subset of L=2 or L=3 layers can be decoded; here each single layer is nearly as good as a dedicated single transmission of the same bandwidth, while combining layers improves quality, though with an expected loss relative to successive refinement because the layers carry redundant information. Against digital baselines built from JPEG2000 or BPG followed by an ideal capacity-achieving channel code, the learned scheme is better or comparable, especially at SNR=1 dB and small bandwidth ratios, and it degrades gracefully when the test SNR or channel model differs from training.

Load-bearing premise

The load-bearing premise is that a single jointly trained encoder can split image information across layers so that every prefix or subset of layers decodes nearly as well as a dedicated single-transmission encoder of the same total bandwidth; this is observed empirically on CIFAR-10 and only for two layers on larger images, not guaranteed by the architecture.

Editorial extensions

If this is right

  • Receivers with different bandwidths can decode the same transmission: a base-layer-only receiver gets a useful image while a richer receiver combines more layers for higher quality, without retraining per deployment.
  • The scheme is resilient to SNR mismatch and fading: quality degrades gradually instead of collapsing, because the analog-style learned mapping has no cliff effect.
  • Dividing a fixed bandwidth into many layers, up to L=16 in the paper, costs almost no PSNR, so bandwidth can be allocated dynamically after training.
  • At low SNR and small bandwidth ratios, the learned layered scheme outperforms digital separation based on JPEG2000 or BPG with an ideal capacity-achieving code.
  • In multiple-description mode, any subset of layers can be decoded; individual layers are near single-transmission quality, but joint decoding incurs some loss relative to successive refinement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the shared-encoder premise would be to fix the trained encoder and train decoders for layer subsets on a held-out dataset; if the encoder is truly near-successfully refinable, these decoders should match the paper's numbers, while a large gap would show the effect depends on joint training rather than on the representation itself.
  • The paper's trade-off curves suggest a deployment rule: for multiple descriptions, choose the loss weight by the expected mix of receivers, favoring joint quality when most receivers get all layers and favoring side quality when many get only one layer.
  • If the near-refinability result transfers to higher-resolution or video sources, learned layered transmission could replace progressive image delivery over noisy channels, with receiver quality set by how many layers are collected rather than by a predefined bitrate.
  • The residual architecture's use of simulated channel realizations hints that feeding back the receiver's actual reconstruction could improve refinement layers; a feedback version is a natural comparison point but is not tested in this paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DeepJSCC-l, a family of deep neural joint source-channel coding schemes for progressive image transmission over noisy channels. The system encodes an image into L channel codewords using a single encoder and decodes from any prefix (successive refinement) or any subset (multiple descriptions) of received layers using dedicated decoders, with alternative single-decoder and residual-transmission architectures also explored. The authors report experiments on CIFAR-10 and ImageNet-trained models evaluated on Kodak, comparing against single-layer DeepJSCC and against JPEG2000/BPG with ideal capacity-achieving channel codes. The central claims are that DeepJSCC-l achieves nearly the same end-to-end PSNR as a single transmission of the same total bandwidth, that it can support up to L=16 layers with negligible loss on CIFAR-10, and that it matches or outperforms digital separation-based baselines at low SNR and small bandwidth while retaining graceful degradation.

Significance. If the empirical claims hold, this is a solid and useful contribution to the learned JSCC literature: it is, to my knowledge, the first practical multiple-description JSCC scheme for real images, and it provides systematic evidence that a shared encoder can learn nearly successively refinable representations over Gaussian channels. The paper's strengths are its clear formulation of the two problems, the comparison of three architectures with different complexity trade-offs, the inclusion of AWGN and Rayleigh fading channels, and the use of ideal separation-based digital baselines as upper bounds. The claims are empirical rather than supported by proofs or guarantees, and no code is released, so reproducibility rests on the completeness of the experimental description. The main weaknesses are that the 'negligible loss' wording is not quantified with numerical gaps or confidence intervals, and the high-layer-count and high-resolution evidence is only partially provided.

major comments (3)
  1. [IV-A, Fig. 4; IV-D, Fig. 7] The central claim that layering causes 'negligible' loss is stated only qualitatively. The text near Fig. 4 says the loss is negligible, and near Fig. 7 that there is 'almost no loss' for up to L=16, but no numerical gap (e.g., average or worst-case PSNR difference from the single-transmission baseline) is reported, and no variance, confidence interval, or multiple-seed results are given. Because the comparison is between two trained models whose difference may be only a few tenths of a dB, the authors should quantify the gap and its uncertainty, for example by reporting the mean and standard error over test images and channel realizations, or over training seeds. Without this, the 'negligible' claim is not falsifiable.
  2. [IV-C, IV-D, Fig. 8] The evidence for the scalability claim 'justifying the use of as many layers as desired' is incomplete. The L=16 result in Fig. 7 is obtained only on CIFAR-10 (32x32 images), while the ImageNet-trained model evaluated on Kodak (Fig. 8) is tested only for L=2. The shared-encoder formulation in Eqs. (4)-(5) has no architectural or theoretical guarantee that the per-prefix distortion remains close to the single-layer bound as resolution or L grows. Since the abstract and Section IV-C make a general claim of negligible loss, the authors should either provide results for L>2 on ImageNet/Kodak, or explicitly scope the claim to CIFAR-10 for L>2 and to L=2 for higher-resolution images.
  3. [IV-E2, Residual Transmission] The residual transmission architecture relies on the transmitter estimating the receiver's previous reconstruction by averaging over m=10 Monte Carlo channel realizations. The paper reports that this scheme achieves results 'very close' to the jointly trained architectures, but no sensitivity analysis with respect to m is provided. Since this estimator is a heuristic component of one of the three architectures claimed to be equivalent in performance, the authors should either vary m and show that the result is insensitive, or temper the equivalence claim.
minor comments (5)
  1. [Abstract] The word 'multiple-decription' is misspelled; it should be 'multiple-description'.
  2. [III, Eq. (4)] The notation in Eq. (4), where the argmin is written outside a sum over subsets, is mathematically awkward because the same encoder parameter theta must be shared across all terms. Please rewrite as a joint optimization or a multi-objective formulation for clarity.
  3. [IV-E2] There is a typo: 'transmissoin' should be 'transmission'.
  4. [V-A, Fig. 14] The repetition-code baseline is described only briefly. Please clarify whether the same codeword is repeated subject to the per-layer average power constraint in Eq. (1), and how the receiver combines the repeated symbols beyond 'averaging'.
  5. [General] No public code or trained models are provided. A reproducibility statement or a commitment to release the code would strengthen the paper, given that all claims are empirical.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the multi-layer results are measured against separately trained single-layer DeepJSCC baselines and external JPEG2000/BPG digital schemes, with no fitted parameter renamed as a prediction.

full rationale

The paper's central claim that DeepJSCC-l can transmit progressively 'with negligible losses' compared with a single transmission is an empirical, post-training observation. The comparisons in Figs. 4, 7, and 8 plot the PSNR of the trained multi-layer model against separately trained single-layer DeepJSCC models of equal bandwidth, not against quantities derived from the multi-layer model's own parameters. Equations (4), (5), and (6) are training objectives, not predictions; they define what is optimized, and the paper's contribution is the observed generalization, which is then tested on held-out CIFAR-10, Kodak, and ImageNet data. The architectures are inherited from the authors' prior work [17] and [19], and those citations are load-bearing as components, but they are published, independently reproducible methods, and the current paper's novelty lies in the multi-layer training formulation rather than in re-deriving those components. Self-citations appear, but they do not substitute for evidence: the 'negligible loss' claim is backed by direct numerical comparison, and the digital baselines (JPEG2000+capacity, BPG+capacity) are external standards. No uniqueness theorem is imported from the authors' prior work, and no fitted parameter is renamed as a prediction. Possible concerns about statistical strength—no error bars, limited L for high-resolution data—are evidence-quality issues, not circularity. The derivation chain is therefore self-contained with respect to its own claims.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The layered code is a learned representation, not a postulated mechanism. The listed free parameters are design choices; the network weights themselves are trained parameters and are treated as implementation details outside this ledger.

free parameters (5)
  • Total bandwidth ratio k/n = c/48 = c = 4, 8, 12, 16 giving 1/12, 1/6, 1/4, 1/3
    Encoder output depth c sets total channel bandwidth; chosen by hand per experiment rather than derived.
  • Layer count L = 1, 2, 3, 4, 5, 8, 16 depending on experiment
    Number of layers is a design choice; results show performance depends only weakly on L.
  • Loss weighting lambda_j and alpha1/alpha2 = lambda_j = 1/L in main results; alpha1/alpha2 varied in Appendix A
    Equal weighting is assumed in Eq. (5) and (6); the paper shows limited impact but does not optimize it.
  • Monte Carlo sample count m for residual architecture = m = 10
    Transmitter estimates the receiver's reconstruction by averaging m simulated channel realizations; chosen empirically.
  • Training SNR = 1, 7, 13, 19 dB in Fig. 5
    One model per training SNR; the graceful degradation claim is tested per model.
assumptions (4)
  • domain assumption AWGN and Rayleigh slow fading channel models with circularly symmetric complex Gaussian noise and power constraint (Eq. 1).
    The learned codes are optimized for these channel statistics; other channels would require retraining.
  • domain assumption MSE/PSNR is the operative distortion measure; no perceptual metric is optimized.
    All training and comparison use MSE; conclusions about image quality are tied to PSNR.
  • ad hoc to paper A single feed-forward CNN encoder shared by all layers can allocate information across layers and subsets without significant loss.
    This is the core structural assumption; it is validated only empirically on CIFAR-10 and Kodak, with no formal guarantee.
  • standard math The capacity-achieving code bound for digital separation assumes zero-overhead channel coding at Shannon capacity.
    Used as an upper bound for digital schemes in Section IV-D; acknowledged as infeasible in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bandwidth-Agile Image Transmission with Deep Joint Source-Channel Coding." pith.science (2026). https://pith.science/paper/RRE5NHGF

@misc{pith2026200912480,
  author       = {Pith},
  title        = {Pith review of: Bandwidth-Agile Image Transmission with Deep Joint Source-Channel Coding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRE5NHGF}},
  note         = {Machine review of arXiv:2009.12480}
}
abstract

We propose deep learning based communication methods for adaptive-bandwidth transmission of images over wireless channels. We consider the scenario in which images are transmitted progressively in layers over time or frequency, and such layers can be aggregated by receivers in order to increase the quality of their reconstructions. We investigate two scenarios, one in which the layers are sent sequentially, and incrementally contribute to the refinement of a reconstruction, and another in which the layers are independent and can be retrieved in any order. Those scenarios correspond to the well known problems of \textit{successive refinement} and \textit{multiple descriptions}, respectively, in the context of joint source-channel coding (JSCC). We propose DeepJSCC-$l$, an innovative solution that uses convolutional autoencoders, and present three architectures with different complexity trade-offs. To the best of our knowledge, this is the first practical multiple-description JSCC scheme developed and tested for practical information sources and channels. Numerical results show that DeepJSCC-$l$ can learn to transmit the source progressively with negligible losses in the end-to-end performance compared with a single transmission. Moreover, DeepJSCC-$l$ has comparable performance with state of the art digital progressive transmission schemes in the challenging low signal-to-noise ratio (SNR) and small bandwidth regimes, with the additional advantage of graceful degradation with channel SNR.

Figures

Figures reproduced from arXiv: 2009.12480 by the authors.

Figure 1
Figure 1. Bandwidth-agile JSCC illustrating successive refinement (left) and multiple descriptions (right). Given an [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The encoder and decoder components used in this paper, introduced in [17]. The notation [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. DeepJSCC-l architecture for progressive wireless image transmission with two layers, performing successive refinement. An input image is encoded into layers z1 and z2, each of them transmitted over different realizations of the noisy channel. L = 1 L 1 N X L j=1 X N i=1 d(x i , xˆ i j ), (5) where d(x i , xˆ i j ) is the MSE distortion between the original image x i and its reconstruction at decoder j, xˆ i j , for … view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: DeepJSCC-l performance for successive refinement with L = 2 layers over a wide range of SNRs, for k1/n = k2/n = 1/12. Colored curves show the performance of reconstructions using both subsets of channel outputs (xˆ1 and xˆ2). Black dashed lines plot the performance of …
Figure 5
Figure 5. Figure 5: DeepJSCC-l performance on successive refinement when there is disparity between training and test channel conditions, typical from multi-user communication. Each color represents the performance over a range of SNR for a DeepJSCC-l model trained for a specific SNR; tri…
Figure 6
Figure 6. Figure 6: (a) Performance of DeepJSCC-l using L = 5 layers over different SNRs. Note that the increase in performance with each refinement layer gradually decreases. (b) Performance of the two first layers (xˆ1 and xˆ2) for DeepJSCC-l trained with different values of L. Note tha…
Figure 7
Figure 7. Figure 7: PSNR vs. bandwidth ratio comparison for L = 1, 2, 4, 8 and 16 layers at (a) SNR = 1dB and (b) SNR = 19dB. DeepJSCC-l presents superior performance for the first layers when compared to a separation-based scheme using JPEG2000 (with 16 layers) or BPG for compression, an…
Figure 9
Figure 9. Figure 9: Although these more complex network architectures produce remarkable performance [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 8
Figure 8. Figure 8: Performance results with encoder/decoder architecture introduced in [19]. Model was trained on ImageNet [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Examples of reconstructions of successive refinement model. Note how flower details are enhanced between [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Performance of alternative successive refinement DeepJSCC- [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Single decoder scheme with two layers. A single decoder is trained with different input sizes, being able to [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Residual transmission scheme with two layers. At each layer, the residual of the previous transmissions is [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: DeepJSCC-l for multiple descriptions problem, where all possible subsets of channel outputs are received and decoded by decoders. Here, with L = 2, decoders 012 and 102 reconstruct the image using distinct sets of channel outputs, while decoder 112 uses all available …
Figure 13
Figure 13. Figure 13: While decoders 012 and 102, each decodes the underlying image from only one of the layers, decoder 112 decodes the same image using both layers. In general, all possible subsets can be indexed with binary numbers formed by L bits, so that the i-th least significant bi…
Figure 14
Figure 14. Figure 14: Performance of multiple description problem on CIFAR-10 test images. (a) [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Examples of reconstructions for different subsets of multiple description problem for [PITH_FULL_IMAGE:figures/full_fig_p026_15.png]
Figure 16
Figure 16. Figure 16: Performance of the single decoder architecture, with the same configurations as in Fig. 14. [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Trade-off between the PSNR achieved by the base layer and that is achieved by combining both layers in [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]
Figure 18
Figure 18. Figure 18: Performance impact of varying the weights of different components in the multiple description problem. (a) [PITH_FULL_IMAGE:figures/full_fig_p029_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    Multiple description coding: compression meets the network,

    V. K. Goyal, “Multiple description coding: compression meets the network,” IEEE Signal Processing Magazine , vol. 18, no. 5, pp. 74–93, 2001

    work page 2001

  2. [2]

    Source coding for a simple network,

    R. M. Gray and A. D. Wyner, “Source coding for a simple network,” Bell System Technical Journal , vol. 53, no. 9, pp. 1681–1721, 1974

    work page 1974

  3. [3]

    Source coding for multiple descriptions,

    J. K. Wolf, A. D. Wyner, and J. Ziv, “Source coding for multiple descriptions,” The Bell System Technical Journal, vol. 59, no. 8, pp. 1417–1426, 1980

    work page 1980

  4. [4]

    Achievable rates for multiple descriptions,

    A. E. Gamal and T. Cover, “Achievable rates for multiple descriptions,” IEEE Transactions on Information Theory, vol. 28, no. 6, pp. 851–857, 1982

    work page 1982

  5. [5]

    Hierarchical coding of discrete sources,

    V. N. Koshelev, “Hierarchical coding of discrete sources,” Problemy peredachi informatsii , vol. 16, no. 3, pp. 31–49, 1980

    work page 1980

  6. [6]

    Successive refinement of information,

    W. H. R. Equitz and T. M. Cover, “Successive refinement of information,” IEEE Transactions on Information Theory, vol. 37, no. 2, pp. 269–275, Mar. 1991

    work page 1991

  7. [7]

    Successive refinement of information: characterization of the achievable rates,

    B. Rimoldi, “Successive refinement of information: characterization of the achievable rates,” IEEE Transactions on Information Theory , vol. 40, no. 1, pp. 253–259, 1994

    work page 1994

  8. [8]

    Successive refinement of vector sources under individual distortion criteria,

    J. Nayak, E. Tuncel, D. Gunduz, and E. Erkip, “Successive refinement of vector sources under individual distortion criteria,” IEEE Transactions on Information Theory , vol. 56, no. 4, pp. 1769–1781, 2010

    work page 2010

Show all 45 references
  1. [9]

    On a source-coding problem with two channels and three receivers,

    L. Ozarow, “On a source-coding problem with two channels and three receivers,” Bell System Technical Journal , vol. 59, no. 10, pp. 1909–1921, 1980

    1909

  2. [10]

    Design of multiple description scalar quantizers,

    V. A. Vaishampayan, “Design of multiple description scalar quantizers,” IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 821–834, 1993

    1993

  3. [11]

    Multiple-description vector quantization with lattice codebooks: design and analysis,

    V. A. Vaishampayan, N. J. A. Sloane, and S. D. Servetto, “Multiple-description vector quantization with lattice codebooks: design and analysis,” IEEE Transactions on Information Theory , vol. 47, no. 5, pp. 1718–1734, 2001

    2001

  4. [12]

    Dithered lattice-based quantizers for multiple descriptions,

    Y. Frank-Dayan and R. Zamir, “Dithered lattice-based quantizers for multiple descriptions,” IEEE Transactions on Information Theory , vol. 48, no. 1, pp. 192–204, 2002. 31

    2002

  5. [13]

    Multiple description trellis-coded quantization,

    H. Jafarkhani and V. Tarokh, “Multiple description trellis-coded quantization,” IEEE Transactions on Com- munications, vol. 47, no. 6, pp. 799–803, 1999

    1999

  6. [14]

    Multiple description coding using pairwise correlating transforms,

    Y. Wang, M. T. Orchard, V. Vaishampayan, and A. R. Reibman, “Multiple description coding using pairwise correlating transforms,” IEEE Transactions on Image Processing , vol. 10, no. 3, pp. 351–366, 2001

    2001

  7. [15]

    On hierarchical joint source-channel coding with degraded side information,

    Y. Steinberg and N. Merhav, “On hierarchical joint source-channel coding with degraded side information,” IEEE Transactions on Information Theory , vol. 52, no. 3, pp. 886–903, 2006

    2006

  8. [16]

    To code or not to code,

    M. Gastpar, “To code or not to code,” p. 170, 2002. [Online]. Available: http://infoscience.epfl.ch/record/33163

    2002

  9. [17]

    Deep joint source-channel coding for wireless image transmission,

    E. Bourtsoulatze, D. Burth Kurka, and D. G¨ und¨ uz, “Deep joint source-channel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 567–579, Sep. 2019

    2019

  10. [18]

    Deep joint source-channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G¨ und¨ uz, “Deep joint source-channel coding for wireless image transmission,” in ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2019, pp. 4774–4778

    2019

  11. [19]

    Joint source-channel coding of images with (not very) deep learning,

    D. Burth Kurka and D. G¨ und¨ uz, “Joint source-channel coding of images with (not very) deep learning,” in Proc of International Zurich Seminar on Information and Communication (IZS 2020) . ETH Zurich, 2020, pp. 90–94

    2020

  12. [20]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT Press, 2016

    2016

  13. [21]

    Deep learning for joint source-channel coding of text,

    N. Farsad, M. Rao, and A. Goldsmith, “Deep learning for joint source-channel coding of text,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , Apr. 2018

    2018

  14. [22]

    Joint source channel coding of information sources using neural networks,

    D. Gunduz, “Joint source channel coding of information sources using neural networks,” Aug. 2018, Patent GB2576702A;WO2020035684A1

    2018

  15. [23]

    NECST: neural joint source-channel coding,

    K. Choi, K. Tatwawadi, T. Weissman, and S. Ermon, “NECST: neural joint source-channel coding,” CoRR, vol. abs/1811.07557, 2018. [Online]. Available: http://arxiv.org/abs/1811.07557

    2018 arXiv

  16. [24]

    M to 1 joint source-channel coding of gaussian sources via dichotomy of the input space based on deep learning,

    Y. M. Saidutta, A. Abdi, and F. Fekri, “M to 1 joint source-channel coding of gaussian sources via dichotomy of the input space based on deep learning,” in 2019 Data Compression Conference (DCC) , 2019, pp. 488–497

    2019

  17. [25]

    Joint source-channel coding for gaussian sources over awgn channels using variational autoencoders,

    ——, “Joint source-channel coding for gaussian sources over awgn channels using variational autoencoders,” in 2019 IEEE International Symposium on Information Theory (ISIT) , 2019, pp. 1327–1331

    2019

  18. [26]

    Analog joint source-channel coding for Gaussian sources over AWGN channels with deep learning,

    Z. Xuan and K. Narayanan, “Analog joint source-channel coding for Gaussian sources over AWGN channels with deep learning,” in 2020 International Conference on Signal Processing and Communications (SPCOM) , 2020, pp. 1–5

    2020

  19. [27]

    Neural network coding,

    L. Liu, A. Solomon, S. Salamatian, and M. M´ edard, “Neural network coding,” in ICC 2020 - 2020 IEEE International Conference on Communications (ICC) , 2020, pp. 1–6

    2020

  20. [28]

    CNN-based analog CSI feedback in FDD MIMO-OFDM systems,

    M. B. Mashhadi, Q. Yang, and D. G¨ und¨ uz, “CNN-based analog CSI feedback in FDD MIMO-OFDM systems,” in IEEE Int’l Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2020, pp. 8579–8583

    2020

  21. [29]

    Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems,

    J. Shao and J. Zhang, “Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems,” in 2020 IEEE International Conference on Communications Workshops (ICC Workshops) , 2020, pp. 1–6

    2020

  22. [30]

    Wireless image retrieval at the edge,

    M. Jankowski, D. G¨ und¨ uz, and K. Mikolajczyk, “Wireless image retrieval at the edge,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 89–100, 2021

    2021

  23. [31]

    Variable rate image compression with recurrent neural networks,

    G. Toderici, S. M. O’Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, “Variable rate image compression with recurrent neural networks,” Proc. of the Int. Conf. on Learning Representations (ICLR), 2016. 32

    2016

  24. [32]

    Real-time adaptive image compression,

    O. Rippel and L. Bourdev, “Real-time adaptive image compression,” in Proc. Int. Conf. on Machine Learning (ICML), vol. 70, Aug. 2017, pp. 2922–2930

    2017

  25. [33]

    Variational image compression with a scale hyperprior,

    J. Ball´ e, D. Minnen, S. Singh, S. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” in Proc. of Int. Conf. on Learning Representations (ICLR) , 2018

    2018

  26. [34]

    Joint autoregressive and hierarchical priors for learned image compression,

    D. Minnen, J. Ball´ e, and G. D. Toderici, “Joint autoregressive and hierarchical priors for learned image compression,” in Advances in Neural Information Processing Systems 31 , S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds. Curran Ass...

    2018

  27. [35]

    Channel-wise autoregressive entropy models for learned image compression,

    D. Minnen and S. Singh, “Channel-wise autoregressive entropy models for learned image compression,” in Int. Conf. on Image Compression (ICIP), 2020 , 2020

    2020

  28. [36]

    Deep multiple description coding by learning scalar quantization,

    L. Zhao, H. Bai, A. Wang, and Y. Zhao, “Deep multiple description coding by learning scalar quantization,” in 2019 Data Compression Conference (DCC) , 2019, pp. 615–615

    2019

  29. [37]

    Multiple description convolutional neural networks for image compression,

    ——, “Multiple description convolutional neural networks for image compression,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 29, no. 8, pp. 2494–2508, 2019

    2019

  30. [38]

    Learning a deep vector quantization network for image compression,

    X. Lu, H. Wang, W. Dong, F. Wu, Z. Zheng, and G. Shi, “Learning a deep vector quantization network for image compression,” IEEE Access, vol. 7, pp. 118 815–118 825, 2019

    2019

  31. [39]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification,” arXiv:1502.01852v1 [cs.CV] , 2015

    2015 arXiv

  32. [40]

    Tensorflow: A system for large-scale machine learning,

    M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving, M. Isard et al., “Tensorflow: A system for large-scale machine learning,” in 12th USENIX Symposium on Operating Systems Design and Implementation , 2016, pp. 265–283

    2016

  33. [41]

    Adam: a method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: a method for stochastic optimization,” in International Conference on Learning Representations (ICLR), 2015

    2015

  34. [42]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” University of Toronto, Tech. Rep., 2009

    2009

  35. [43]

    The zero error capacity of a noisy channel,

    C. Shannon, “The zero error capacity of a noisy channel,” IRE Transactions on Information Theory , vol. 2, no. 3, pp. 8–19, Sep. 1956

    1956

  36. [44]

    Density modeling of images using a generalized normalization transformation,

    J. Ball´ e, V. Laparra, and E. P. Simoncelli, “Density modeling of images using a generalized normalization transformation,” arXiv preprint arXiv:1511.06281 , 2015

    2015 arXiv

  37. [45]

    DeepJSCC-f: Deep joint source-channel coding of images with feedback,

    D. B. Kurka and D. G¨ und¨ uz, “DeepJSCC-f: Deep joint source-channel coding of images with feedback,”IEEE Journal on Selected Areas in Information Theory , vol. 1, no. 1, pp. 178–193, May 2020

    2020

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.