pith. sign in

arxiv: 2606.29578 · v1 · pith:J6TXZ4JDnew · submitted 2026-06-28 · 💻 cs.IT · eess.SP· math.IT

SoftBinary Coding: A New Information-Theoretic Neural Compression Paradigm

Ezgi Ozyilkan , Sharang M. Sriramu , Elza Erkip , Aaron B. Wagner , Jona Ball\'e This is my paper

Pith reviewed 2026-06-30 01:51 UTC · model grok-4.3

classification 💻 cs.IT eess.SPmath.IT
keywords neural compressionbinary latent spacechannel simulationrate-distortionvector quantizationSoftBinary Codinginformation theorydiscrete representations
0
0 comments X

The pith

SoftBinary Coding reaches optimal rate-distortion bounds in neural compression through stochastic binary latents and a rate-optimal channel simulation scheme.

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

Nonlinear Transform Coding relies on continuous real-valued latents but incurs train-test mismatch from quantization, a smoothness bias that blocks optimality for some sources, and reduced shaping gain from complex vector quantization. SoftBinary Coding replaces this with a stochastic binary latent space that uses discrete representations. These representations are compressed by a novel fast binary channel simulation scheme, which the paper proves is rate-optimal. Experiments on information-theoretic sources show performance gains that close NTC limitations, while vector quantization tests on i.i.d. sources yield state-of-the-art results that surpass Trellis Coded Quantization for the Gaussian source.

Core claim

SBC employs discrete representations and compresses them through a novel fast binary channel simulation scheme, for which we provide a proof of rate optimality. Experimental gains on information-theoretic sources provide both theoretical and practical closure to NTC's limitations, establishing discrete binary structures as a viable path toward reaching optimal rate--distortion bounds. Surprisingly, SBC also achieves state-of-the-art performance on vector quantization of i.i.d. sources, exceeding Trellis Coded Quantization of the Gaussian source.

What carries the argument

stochastic binary latent space with fast binary channel simulation scheme, which produces discrete representations and enables rate-optimal compression.

If this is right

  • Neural compression can avoid train-test mismatch and smoothness bias by switching to discrete binary latents.
  • The channel simulation scheme achieves rate optimality, allowing the system to reach theoretical rate-distortion bounds.
  • SBC delivers measurable gains on information-theoretic sources and sets new performance records for vector quantization of i.i.d. sources.
  • Discrete binary structures serve as a practical alternative to continuous transforms for optimal compression.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the binary latent approach extends to image or video data, it could serve as a drop-in replacement for continuous NTC pipelines in deployed codecs.
  • The rate-optimality proof for the simulation scheme may generalize to other discrete alphabets beyond binary.
  • Hybrid systems that combine SBC latents with existing learned transforms could improve compression for sources where NTC currently underperforms.

Load-bearing premise

The stochastic binary latent space combined with the channel simulation scheme can be trained end-to-end without reintroducing train-test mismatch or smoothness bias.

What would settle it

Run SBC on a standard Gaussian source at a fixed rate and measure whether the achieved distortion meets or beats the known rate-distortion bound while the simulation rate stays at or below the source entropy.

Figures

Figures reproduced from arXiv: 2606.29578 by Aaron B. Wagner, Elza Erkip, Ezgi Ozyilkan, Jona Ball\'e, Sharang M. Sriramu.

Figure 1
Figure 1. Figure 1: Training scheme for learning-based lossy neural compression with channel simulation. The neural network fθ produces parameters V of the encoder distribution, a parametric family F(Z | V ). The latent representation Z is a sample from this family (operationally, channel simulation produces the sample at the decoder). The encoder qθ and prior pψ (a model of the marginal distribution of Z) evaluated over Z de… view at source ↗
Figure 2
Figure 2. Figure 2: System diagram for the operational scheme using channel simulation (Sec. 3). We concatenate the encoded messages V L (i) = fθ(Xi) for several independent source realizations and use Algorithms 1 and 2 to generate the latent samples Z N at the decoder. Algorithm 1 Generalized PolarSim: Encoder side Input: Block length N Input: Source sequence v N ∼ QN i=1 PVi Input: Random seed s N i.i.d. ∼ Unif(0, 1) Input… view at source ↗
Figure 3
Figure 3. Figure 3: Rate–distortion performance on the circle. One-shot entropy–distortion (E-D) bound is due to Bhadane et al. (2022). 3 4 5 6 7 8 −30 −27 −24 −21 −18 better rate [bits] distortion [dB] SBC L = 8 SBC L = 4 NTC one-shot E-D [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Rate–distortion performance on the ramp. One-shot entropy–distortion (E-D) bound is due to Bhadane et al. (2022). It is worth highlighting our use of a larger latent dimension L for SBC (4, 8, or 32 across experiments) compared to NTC (L = 1 in most cases). This follows directly from the two schemes’ capacity constraints: each SBC latent dimension corresponds to a single bit, capping the rate at L bits per… view at source ↗
Figure 6
Figure 6. Figure 6: Rate–distortion performance on distributed compression of Y = X + N with X ∼ N (0, 1) and N ∼ N (0, 10−1 ), where Y is side information (NTC: L = 1; SBC: L = 32). For DIS￾CUS by Pradhan & Ramchandran (2003), we include data points obtained with trellis-based quantization and coset construction, available at R ∈ {1, 2} bits. The asymptotic rate–distortion bound (R-D) with side information, due to Wyner & Zi… view at source ↗
Figure 8
Figure 8. Figure 8: Rate–distortion performance on i.i.d. uniform source (NTC: L = 1; SBC: L = 32). Rate–distortion (R-D) points at R = {1, 2, 3} bits and TCQ points with 256 states are obtained from Taubman & Marcellin (2013). tizers. This currently limits our experiments to low￾dimensional sources. However, we believe that in particular for the kind of low-rate applications that NTC is now in￾creasingly being used for, such… view at source ↗
read the original abstract

Neural compression is currently dominated by Nonlinear Transform Coding (NTC), which maps data to real-valued latents via continuous transforms. Despite its success, NTC suffers from train-test mismatch due to non-differentiable quantization, a ``smoothness bias" inherent in continuous transforms that precludes optimality for certain sources, and a loss of ``shaping gain" due to the complexity of including high-dimensional vector quantization. We propose SoftBinary Coding (SBC), an end-to-end learning paradigm that bypasses these limitations by using a stochastic binary latent space. In the spirit of vector quantization, SBC employs discrete representations and compresses them through a novel fast binary channel simulation scheme, for which we provide a proof of rate optimality. Experimental gains on information-theoretic sources provide both theoretical and practical closure to NTC's limitations, establishing discrete binary structures as a viable path toward reaching optimal rate--distortion bounds. Surprisingly, SBC also achieves state-of-the-art performance on vector quantization of i.i.d. sources, exceeding Trellis Coded Quantization of the Gaussian source.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes SoftBinary Coding (SBC), an end-to-end neural compression paradigm that replaces the continuous latents of Nonlinear Transform Coding (NTC) with a stochastic binary latent space. Data are mapped to discrete binary representations that are then compressed via a novel fast binary channel simulation scheme; the authors provide a proof of rate optimality for this scheme. The work claims that SBC eliminates NTC's train-test mismatch, smoothness bias, and loss of shaping gain, delivers experimental gains on information-theoretic sources, and achieves state-of-the-art vector-quantization performance on i.i.d. sources, surpassing Trellis Coded Quantization on the Gaussian source.

Significance. If the rate-optimality guarantee survives composition with the learned transform and the reported gains are reproducible, the result would establish discrete binary structures as a practical route to optimal rate-distortion performance, directly addressing three long-standing limitations of the dominant NTC framework.

major comments (2)
  1. [Proof of rate optimality (binary channel simulation scheme)] The abstract asserts a proof of rate optimality for the fast binary channel simulation scheme, yet the central claim (closure of NTC limitations and attainment of optimal bounds) requires this optimality to hold after the scheme is embedded inside an end-to-end learned transform. The manuscript must explicitly state whether the proof assumes i.i.d. binary inputs or a memoryless channel independent of the preceding transform; if dependencies introduced by the learned mapping alter the effective channel, the optimality guarantee does not automatically transfer.
  2. [Experimental results] Experimental claims of superiority and SOTA vector-quantization performance are presented without dataset descriptions, error bars, or training details in the abstract; the full manuscript must supply these controls so that the reported gains can be verified against the information-theoretic sources and the Gaussian VQ benchmark.
minor comments (2)
  1. [Method] Notation for the stochastic binary latent space and the channel simulation parameters should be introduced with explicit definitions before the proof is presented.
  2. [Experiments] The abstract states that SBC 'achieves state-of-the-art performance on vector quantization of i.i.d. sources'; the corresponding table or figure should report the exact rate-distortion points and the baseline Trellis Coded Quantization implementation used for comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate where revisions will be made.

read point-by-point responses

  1. Referee: [Proof of rate optimality (binary channel simulation scheme)] The abstract asserts a proof of rate optimality for the fast binary channel simulation scheme, yet the central claim (closure of NTC limitations and attainment of optimal bounds) requires this optimality to hold after the scheme is embedded inside an end-to-end learned transform. The manuscript must explicitly state whether the proof assumes i.i.d. binary inputs or a memoryless channel independent of the preceding transform; if dependencies introduced by the learned mapping alter the effective channel, the optimality guarantee does not automatically transfer.

    Authors: We thank the referee for this important observation. The provided proof establishes rate optimality of the fast binary channel simulation under the standard assumptions of i.i.d. binary inputs and a memoryless channel. In the SBC architecture the learned transform produces the binary latents that are subsequently passed to this simulation module; the end-to-end training therefore optimizes the overall rate-distortion trade-off while the simulation step itself remains rate-optimal for any binary sequence it receives. We will revise the manuscript to state these assumptions explicitly and to clarify that dependencies, if present, are absorbed into the learned transform parameters rather than violating the simulation guarantee. revision: yes

  2. Referee: [Experimental results] Experimental claims of superiority and SOTA vector-quantization performance are presented without dataset descriptions, error bars, or training details in the abstract; the full manuscript must supply these controls so that the reported gains can be verified against the information-theoretic sources and the Gaussian VQ benchmark.

    Authors: Space limitations prevent inclusion of these details in the abstract. The full manuscript already contains a dedicated experimental section that describes the information-theoretic sources, the Gaussian VQ benchmark, training procedures, and reports results with error bars obtained from multiple independent runs. These controls are provided to enable verification of the claimed gains. We will ensure the presentation is sufficiently prominent and, if the referee deems any aspect still insufficient, we are prepared to expand it. revision: no

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The provided abstract and text present the rate-optimality result as a separate proof for the binary channel simulation scheme, independent of the learned transform. No equations or claims reduce the central results (rate-distortion closure, experimental gains) to inputs by construction, self-definition, or fitted parameters renamed as predictions. No self-citation load-bearing steps or ansatz smuggling are visible. The derivation chain is self-contained, relying on an external proof and experiments rather than circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities can be extracted. The stochastic binary latent space and channel simulation scheme are introduced as new but their internal assumptions are not detailed.

pith-pipeline@v0.9.1-grok · 5733 in / 1137 out tokens · 40257 ms · 2026-06-30T01:51:33.029514+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  2. [2]

    Integer Networks for Data Compression with Latent-Variable Models , year =

    Ballé, Jona and Minnen, David and Johnston, Nick , booktitle =. Integer Networks for Data Compression with Latent-Variable Models , year =. S1zz2i0cY7 , eprinttype =

  3. [3]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  4. [4]

    M. J. Kearns , title =

  5. [5]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  6. [6]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  7. [7]

    Suppressed for Anonymity , author=

  8. [8]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  9. [9]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  10. [10]

    Ozyilkan, Ezgi and Ballé, Jona and Erkip, Elza , booktitle=. Learned. 2023 , volume=

    2023

  11. [11]

    Neural Distributed Compressor Discovers Binning , year=

    Ozyilkan, Ezgi and Ballé, Jona and Erkip, Elza , journal=. Neural Distributed Compressor Discovers Binning , year=

  12. [12]

    and Fischer, T.R

    Marcellin, M.W. and Fischer, T.R. , journal=. Trellis coded quantization of memoryless and Gauss-Markov sources , year=

  13. [13]

    The Thirteenth International Conference on Learning Representations , year=

    Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding , author=. The Thirteenth International Conference on Learning Representations , year=

  14. [14]

    International Conference on Learning Representations , year=

    End-to-end Optimized Image Compression , author=. International Conference on Learning Representations , year=

  15. [15]

    2020 , booktitle =

    Agustsson, Eirikur and Theis, Lucas , title =. 2020 , booktitle =

    2020

  16. [16]

    International Conference on Learning Representations , year=

    Variational image compression with a scale hyperprior , author=. International Conference on Learning Representations , year=

  17. [17]

    and Minnen, David and Singh, Saurabh and Johnston, Nick and Agustsson, Eirikur and Hwang, Sung Jin and Toderici, George , journal=

    Ballé, Johannes and Chou, Philip A. and Minnen, David and Singh, Saurabh and Johnston, Nick and Agustsson, Eirikur and Hwang, Sung Jin and Toderici, George , journal=. Nonlinear Transform Coding , year=

  18. [18]

    Workshop on Machine Learning and Compression, NeurIPS 2024 , year=

    Breaking Smoothness: The Struggles of Neural Compressors with Discontinuous Mappings , author=. Workshop on Machine Learning and Compression, NeurIPS 2024 , year=

    2024

  19. [19]

    and Ziv, J

    Wyner, A. and Ziv, J. , journal=. The Rate--Distortion Function for Source Coding with Side Information at the Decoder , year=

  20. [20]

    and Thomas, Joy A

    Cover, Thomas M. and Thomas, Joy A. , title =. 2006 , isbn =

    2006

  21. [21]

    and Ramchandran, K

    Pradhan, S.S. and Ramchandran, K. , journal=. Distributed source coding using syndromes (DISCUS): design and construction , year=

  22. [22]

    Advances in Neural Information Processing Systems , volume =

    Richter, Lorenz and Boustati, Ayman and N\". Advances in Neural Information Processing Systems , volume =

  23. [23]

    International Conference on Learning Representations (ICLR) , year =

    Auto-Encoding Variational Bayes , author =. International Conference on Learning Representations (ICLR) , year =

  24. [24]

    and Ballé, Johannes , booktitle=

    Bhadane, Sourbh and Wagner, Aaron B. and Ballé, Johannes , booktitle=. Do Neural Networks Compress Manifolds Optimally? , year=

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Ball\'e, Jona and Versari, Luca and Dupont, Emilien and Kim, Hyunjik and Bauer, Matthias , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  26. [26]

    Advances in Neural Information Processing Systems , volume =

    High-Fidelity Generative Image Compression , author =. Advances in Neural Information Processing Systems , volume =

  27. [27]

    Gray, R. M. , journal=. Vector quantization , year=

  28. [28]

    , journal=

    Ziv, J. , journal=. On universal quantization , year=

  29. [29]

    , journal=

    Schuchman, L. , journal=. Dither Signals and Their Effect on Quantization Noise , year=

  30. [30]

    IRE Transactions on Information Theory , volume =

    Picture coding using pseudo-random noise , author =. IRE Transactions on Information Theory , volume =. 1962 , publisher =

    1962

  31. [31]

    and Feder, M

    Zamir, R. and Feder, M. , journal=. On universal quantization by randomized uniform/lattice quantizers , year=

  32. [32]

    Proceedings of the 34th International Conference on Machine Learning , pages =

    Real-Time Adaptive Image Compression , author =. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , editor =

    2017

  33. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Joint Autoregressive and Hierarchical Priors for Learned Image Compression , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  34. [34]

    Theis and W

    L. Theis and W. Shi and A. Cunningham and F. Huszár. Lossy Image Compression with Compressive Autoencoders. International Conference on Learning Representations

  35. [35]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Fast Channel Simulation via Error-Correcting Codes , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  36. [36]

    Channel Polarization: A Method for Constructing Capacity-Achieving Codes for Symmetric Binary-Input Memoryless Channels , year=

    Arikan, Erdal , journal=. Channel Polarization: A Method for Constructing Capacity-Achieving Codes for Symmetric Binary-Input Memoryless Channels , year=

  37. [37]

    , journal =

    Rissanen, Jorma J. , journal =. Generalized. 1976 , publisher =

    1976

  38. [38]

    1999 , publisher =

    Sphere Packings, Lattices and Groups , author =. 1999 , publisher =

    1999

  39. [39]

    2013 , publisher =

    JPEG2000 Image Compression Fundamentals, Standards and Practice , author =. 2013 , publisher =

    2013

  40. [40]

    2014 , isbn =

    Zamir, Ram , title =. 2014 , isbn =

    2014

  41. [41]

    and Ballé, Jona , booktitle=

    Wagner, Aaron B. and Ballé, Jona , booktitle=. Neural Networks Optimally Compress the Sawbridge , year=

  42. [42]

    2022 , isbn =

    Li, Jiahao and Li, Bin and Lu, Yan , title =. 2022 , isbn =. doi:10.1145/3503161.3547845 , booktitle =

    work page doi:10.1145/3503161.3547845 2022

  43. [43]

    Kot and Bihan Wen , booktitle=

    Yufei Wang and Zhihao Li and Lanqing Guo and Wenhan Yang and Alex C. Kot and Bihan Wen , booktitle=. Context. 2024 , url=

    2024

  44. [44]

    Learning Convolutional Transforms for Lossy Point Cloud Geometry Compression , year=

    Quach, Maurice and Valenzise, Giuseppe and Dufaux, Frederic , booktitle=. Learning Convolutional Transforms for Lossy Point Cloud Geometry Compression , year=

  45. [45]

    and Bourdev, Lubomir , title =

    Rippel, Oren and Nair, Sanjay and Lew, Carissa and Branson, Steve and Anderson, Alexander G. and Bourdev, Lubomir , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

  46. [46]

    Strong Functional Representation Lemma and Applications to Coding Theorems , year=

    Li, Cheuk Ting and El Gamal, Abbas , journal=. Strong Functional Representation Lemma and Applications to Coding Theorems , year=

  47. [47]

    2024 , volume =

    Foundations and Trends® in Communications and Information Theory , title =. 2024 , volume =. doi:10.1561/0100000141 , issn =

    work page doi:10.1561/0100000141 2024

  48. [48]

    Bounds for entropy and divergence for distributions over a two-element set , journal =

    Tops. Bounds for entropy and divergence for distributions over a two-element set , journal =

  49. [49]

    2010 IEEE International Symposium on Information Theory , pages=

    Source polarization , author=. 2010 IEEE International Symposium on Information Theory , pages=. 2010 , organization=

    2010

  50. [50]

    IEEE Transactions on information Theory , volume=

    Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels , author=. IEEE Transactions on information Theory , volume=. 2009 , publisher=

    2009

  51. [51]

    IBM Journal of research and development , volume=

    Arithmetic coding , author=. IBM Journal of research and development , volume=. 1979 , publisher=

    1979

  52. [52]

    Inequalities of

    Kadelburg, Zoran and Dukic, Dusan and Lukic, Milivoje and Matic, Ivan , journal=. Inequalities of. 2005 , publisher=

    2005

  53. [53]

    Advances in Neural Information Processing Systems , volume=

    Greedy Poisson rejection sampling , author=. Advances in Neural Information Processing Systems , volume=

  54. [54]

    European Conference on Computer Vision , year=

    HAC: Hash-grid Assisted Context for 3D Gaussian Splatting Compression , author=. European Conference on Computer Vision , year=

  55. [55]

    Proceedings of the 1st International Workshop on Advances in Point Cloud Compression, Processing and Analysis , year =

    Pang, Jiahao and Lodhi, Muhammad Asad and Tian, Dong , title =. Proceedings of the 1st International Workshop on Advances in Point Cloud Compression, Processing and Analysis , year =. doi:10.1145/3552457.3555727 , isbn =

    work page doi:10.1145/3552457.3555727

  56. [56]

    Foundations and Trends

    Polarization and polar codes , author=. Foundations and Trends. 2012 , publisher=

    2012

  57. [57]

    IEEE Transactions on Information Theory , volume=

    Finite-length scaling for polar codes , author=. IEEE Transactions on Information Theory , volume=. 2014 , publisher=

    2014

  58. [58]

    The art of computer programming , volume=

    Seminumerical algorithms , author=. The art of computer programming , volume=. 1981 , publisher=

    1981

  59. [59]

    IEEE Transactions on Information Theory , volume=

    Empirical and strong coordination via soft covering with polar codes , author=. IEEE Transactions on Information Theory , volume=. 2018 , publisher=

    2018

  60. [60]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Channel Simulation and Distributed Compression with Ensemble Rejection Sampling , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  61. [61]

    Canadian Journal of Mathematics , volume=

    Notes on sphere packings , author=. Canadian Journal of Mathematics , volume=. 1967 , publisher=. doi:10.4153/CJM-1967-017-0 , issn=

    work page doi:10.4153/cjm-1967-017-0 1967

  62. [62]

    Towards Reproducible Learning-Based Compression , year=

    Pang, Jiahao and Lodhi, Muhammad Asad and Ahn, Junghyun and Huang, Yuning and Tian, Dong , booktitle=. Towards Reproducible Learning-Based Compression , year=

  63. [63]

    Rate-Distortion Optimized Post-Training Quantization for Learned Image Compression , year=

    Shi, Junqi and Lu, Ming and Ma, Zhan , journal=. Rate-Distortion Optimized Post-Training Quantization for Learned Image Compression , year=

  64. [64]

    Forney, G. D. , journal=. Trellis shaping , year=

  65. [65]

    Deep Learning-Based Image Compression with Trellis Coded Quantization , year=

    Li, Binglin and Akbari, Mohammad and Liang, Jie and Wang, Yang , booktitle=. Deep Learning-Based Image Compression with Trellis Coded Quantization , year=

  66. [66]

    , booktitle=

    Ballé, Jona and Laparra, Valero and Simoncelli, Eero P. , booktitle=. End-to-end optimization of nonlinear transform codes for perceptual quality , year=

  67. [67]

    Advances in Neural Information Processing Systems 30 (NIPS 2017) , pages=

    Neural Discrete Representation Learning , author=. Advances in Neural Information Processing Systems 30 (NIPS 2017) , pages=

    2017

  68. [68]

    Algorithmic Polarization for Hidden Markov Models

    Algorithmic polarization for hidden markov models , author=. arXiv preprint arXiv:1810.01969 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  69. [69]

    Concatenated Codes. 1965

    1965

  70. [70]

    IEEE Transactions on Information Theory , volume=

    Polar coding for processes with memory , author=. IEEE Transactions on Information Theory , volume=. 2019 , publisher=

    2019

  71. [71]

    List Decoding of Polar Codes , year=

    Tal, Ido and Vardy, Alexander , journal=. List Decoding of Polar Codes , year=

  72. [72]

    Polar Coding for Non-Stationary Channels , year=

    Mahdavifar, Hessam , journal=. Polar Coding for Non-Stationary Channels , year=

  73. [73]

    Categorical Reparameterization with

    Jang, Eric and Gu, Shixiang and Poole, Ben , booktitle=. Categorical Reparameterization with

  74. [74]

    International Conference on Learning Representations (ICLR) , year=

    The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables , author=. International Conference on Learning Representations (ICLR) , year=

  75. [75]

    , journal=

    Lloyd, Stuart P. , journal=. Least squares quantization in