Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Latent-attention Based Transformer for Near ML Polar Decoding in Short-code Regime

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A redesigned transformer decoder reaches near-maximum-likelihood error rates for short polar codes and keeps one architecture viable across code rates and lengths.

desk verdict A genuinely new transformer decoder for short polar codes with a serious label-source ambiguity in its near-ML claim; deserves peer review after major revision. read the letter →

arxiv 2507.14951 v1 pith:YTBANNJG submitted 2025-07-20 eess.SP

Hongzhi Zhu , Wei Xu , Xiaohu You This is my paper
classification eess.SP
keywords PolarcodesTransformerdecoderLatentattentionCode-awaremaskingNearmaximum-likelihooddecodingShort-blocklengthDeeplearningchannelGeneralizationacrosscodeconfigurations
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 seeks to establish that a transformer-based decoder can reach near-maximum-likelihood (ML) bit error and block error rates for short polar codes, a regime where generic transformer decoders have trailed algebraic decoders. The proposed LAT decoder replaces ordinary self-attention with a code-aware latent-attention mechanism that learns Query and Key matrices separately from the Value matrix, encoding frozen-bit constraints and bit positions into the attention weights. A front-padding scheme and a code-aware mask let a single trained architecture adapt to different code lengths and rates, instead of requiring a dedicated network per code configuration. Numerical results report near-ML BER/BLER for (N=16, k=8) and stable performance comparable to SCL decoding across N in {8,16} and rates 0.25–0.75. If the claims hold, deep-learning decoding becomes practical for short-block ultra-reliable links without sacrificing the error-rate optimality that specialized algebraic decoders provide.

What carries the argument

The load-bearing object is the latent-attention block and its code-aware mask. Each layer computes attention weights as $\mathrm{softmax}\left(\frac{Q_{ib}K_{ib}^T}{\sqrt{d_m}}\right)V_{ib}$, where $Q_{ib}=Q_f+P_{ib}$ combines a learned frozen-bit query bias with the latent positional encoding $P_{ib}$, $K_{ib}$ comes from a layer-specific positional key encoder, and $V_{ib}=[V_y,U_f]$ concatenates embedded channel values with the frozen-bit embedding. The code-aware mask zeros out illegal connections among padding, frozen, and message positions, and front-padding preserves absolute bit indices so the same network can shift between code rates and lengths by mask replacement. This mechanism is what carries the near-ML performance and the generalization claim.

What would settle it

Retrain the LAT decoder with the same protocol but with experience-reflow reference labels generated by SC decoding only, then compare BER/BLER at N=16, k=8 against the ML bound. If near-ML error rates persist, the architecture is responsible; if they fall back to SC-level performance, the ML benchmark was supplying the near-ML result through training labels.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the short-code failure of transformer decoders for polar codes is architectural rather than fundamental: generic self-attention computes context correlations that distort the positional, hierarchical dependencies of polarization, and this can be fixed by decoupling $Q$, $K$, and $V$. In the LAT decoder, layer-specific latent positional encoders generate the Key matrix and a positional Query bias, while the Value matrix concatenates channel-value embeddings with frozen-bit embeddings; a code-aware mask enforces that only valid decoding connections contribute to attention. The paper supports this with a score-function view from Tweedie's formula, arguing that the layer-wise residual updates approximate the ML estimator's gradient, and with an energy-based argument that the training objective is optimizable. The reported outcome is BER/BLER close to ML and SCL for N=16, k=8, better than the ECCT transformer baseline, and stable performance across N in {8,16} and rates 0.25–0.75 after configuration-specific fine-tuning.

Load-bearing premise

The load-bearing premise is that the near-ML curves reflect the architecture rather than the benchmark: Eq. (44) allows the ML decoder to supply experience-reflow training labels, and the model is fine-tuned on each test configuration before evaluation; if either supervision channel is removed, the claimed generalization may collapse.

Editorial extensions

If this is right

  • If the claims hold, transformer decoders can operate at ML error-rate limits for short polar codes without exhaustive ML search or large decoding lists.
  • A single LAT architecture with code-aware masks can serve multiple (N, R) configurations, removing the retrain-per-code requirement of ECCT-style transformer decoders.
  • The training recipe—entropy-aware importance sampling, experience reflow, dynamic label smoothing, and a BLER-aware KLD loss—can be reused as a package for other neural decoders.
  • The larger BER gain than BLER gain indicates the continuous KLD objective mainly sharpens bit decisions; further BLER improvement would need a stronger discrete-loss regularizer.

Reading between the lines

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

  • Inference: because the masked attention weights are interpretable step-wise dependencies, the same latent-attention design could be lifted to other structured codes whose decoding constraints can be written as masks; the paper itself tests only polar codes.
  • Inference: the near-ML result may be partly inherited from the benchmark if the experience-reflow labels in Eq. (44) are generated by the ML decoder; retraining with SC-only labels would reveal how much of the performance is architectural.
  • Inference: the generalization claim is fast adaptation rather than zero-shot transfer, since the paper reports degraded performance without per-configuration fine-tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a decoder-only transformer architecture, the latent-attention transformer (LAT), for short-length polar codes. The architecture replaces standard self-attention with a latent-attention mechanism in which Query and Key matrices are generated by layer-specific latent positional encoders, augmented by frozen-bit prior knowledge, a front-padding scheme, and a code-aware mask that adapts to varying code lengths and rates. The training pipeline combines entropy-aware importance sampling, an experience-reflow buffer, dynamic label smoothing, and a BLER-aware regularization term. The authors report that, after per-configuration fine-tuning, the LAT decoder achieves BER and BLER close to those of ML decoding for a (16,8) polar code, and that it adapts to other short-code configurations more flexibly than the ECCT baseline. The central claims are that LAT achieves near-ML performance in the short-code regime and generalizes across code rates and lengths.

Significance. If the near-ML and generalization claims were fully supported, the paper would offer a useful contribution: a transformer decoder that keeps a single architecture across several short polar-code configurations while approaching ML performance, and a training recipe that sharpens performance near the decoding boundary. The paper also contains interesting ideas: the front-padding scheme that preserves absolute bit positions, the code-aware mask that isolates padding and frozen positions, and the layer-specific latent positional encoders. These are concrete architectural proposals that go beyond a direct application of the ECCT. However, the current evidence for the main claims is limited: the near-ML result is shown for one code configuration, the evaluation protocol includes per-configuration fine-tuning before testing, the source of the experience-reflow labels is not specified, and no confidence intervals or multiple-run statistics are reported. The theoretical section does not provide a rigorous substitute for these missing empirical controls. The contribution is therefore promising but not yet established.

major comments (4)
  1. [V-B, Eq. (44)] The paper does not specify which conventional decoder produced the reference labels m* in Eq. (44). Since the ML decoder is explicitly listed as an option and is also the benchmark in Fig. 8, the near-ML curves may be partly fitted to ML decisions rather than independently achieved. Please state exactly which decoder generated m* for the results in Section VI, and provide an ablation in which m* is generated by a strictly weaker decoder (e.g., SC or SCL) to show that the reported performance is not a consequence of label leakage from the ML benchmark.
  2. [VI-B and VI-C] The reported results are obtained after 100 epochs of per-configuration fine-tuning, and Section VI-C explicitly states that performance degrades when no fine-tuning is employed before testing. This means the cross-configuration results in Fig. 9 are not zero-shot generalization but a form of retraining. The phrase "robust generalization capabilities" in the abstract is therefore misleading. Please report the zero-shot (no-fine-tuning) performance for all configurations in Fig. 9, and define precisely what is meant by generalization, since the current protocol conflates adaptation with generalization.
  3. [VI-B, Fig. 8] The central near-ML claim rests on a single code configuration (N=16, k=8) with no confidence intervals, no number of independent training runs, and no number of test blocks. Transformer training is stochastic and the reported margin to ML is small in BLER; without repeated runs and variance estimates, the "near ML" and "outperforms SC by 0.80 dB" statements are not statistically supported. Please provide results over multiple seeds, report mean and spread, and include at least two additional short-code configurations in the ML-benchmark comparison.
  4. [IV-A, Eq. (32)] The mathematical foundation section states E{m|y}=E{x|y}=y+σ²∇log p(y), but this equation is dimensionally and semantically problematic: m has length k, x is a binary codeword of length N, and y+σ²∇log p(y) is the posterior mean of the continuous BPSK signal, not the posterior mean of the bits. This conflation does not directly invalidate the empirical results, but it is presented as a foundation for the architecture and should be corrected or removed.
minor comments (5)
  1. [III-C, Eqs. (21)-(23)] The layer index ib is used for Qib, Kib, and Vib, but Eq. (18) and Eq. (20) use Q, K, V without specifying how the layer-specific matrices are fed into the multi-head attention; please make the index propagation explicit.
  2. [V-B, Eq. (44)] The retention condition \tilde{m}_{i,m_i} I(\hat{m}_i \neq m_i) < p_b mixes a soft probability with a hard indicator; the intended logic of the buffer update should be stated in words or with a clearer formula.
  3. [VI-C, Fig. 9] The text describing Fig. 9 says panels (a) and (b) are BER and BLER across code configurations, while the figure caption lists (a) and (b) as Eb/N0 variation; please align the text with the captions.
  4. [Throughout] There are numerous typos and wording issues, including "addvancing," "is formulates as," "duel input sequences," "the propose latent attention mechanism," and "decoding detectors"; a careful language edit is needed.
  5. [Table I] The symbol Nb is used both for the number of decoding layers in Section III and for the batch size in Table I; this notation clash should be resolved.

Circularity Check

2 steps flagged · score 5.0 of 10

The near-ML claim is partly trained into the model if Eq. (44)'s reference labels came from the ML decoder, and the generalization claim is supported only after per-configuration fine-tuning.

  1. fitted input called prediction [Section V-B, Eq. (44); Section VI-B, Fig. 8]
    "m⋆ denotes a reference decoding label obtained through conventional decoders (e.g., SCL decoder or ML decoder). ... The LAT decoder is trained for additional 100 epochs as fine-tuning for a specific code configuration before test."

    If the ML decoder generated m*, the neural decoder is supervised by ML decisions during experience reflow; reporting near-ML BER/BLER against the ML decoder then largely restates the training target. The paper does not specify which reference decoder produced m* and provides no ablation with a strictly weaker label source, so the central near-ML claim is not independently established.

  2. other [Section VI-C, Fig. 9]
    "Notably, performance degradation occurs when no fine-tuning is employed before testing. Nevertheless, the LAT decoder exhibits rapid adaptability through test-time fine-tuning."

    The claimed generalization across code rates and lengths is evaluated only after fine-tuning the model on each tested configuration. The paper's own admission that performance degrades without fine-tuning means the Fig. 9 curves reflect per-configuration adaptation rather than zero-shot transfer, so the 'robust generalization' claim is fitted to the evaluated configurations rather than predicted.

full rationale

The architectural contributions (latent attention, code-aware mask, front-padding) are self-contained and do not reduce to their inputs, and there is no load-bearing self-citation. The mathematical foundation in Section IV is heuristic and does not smuggle in the result via citation. However, the central empirical claim of near-ML performance is not independently anchored: Section V-B Eq. (44) explicitly permits the ML decoder to be the source of experience-reflow training labels, while Section VI-B benchmarks against the ML decoder. If that option was exercised, the near-ML curves in Fig. 8 partly measure imitation of the ML decoder rather than an independent decoding capability. The paper never states which reference decoder produced m* and offers no ablation with a strictly weaker label source; Section VI-D preserves experience reflow in the weakened baseline, so that comparison does not resolve the ambiguity. Additionally, Section VI-C concedes that performance degrades without fine-tuning, so the cross-configuration generalization claim is demonstrated only after per-configuration fine-tuning. These are support gaps rather than proof that the architecture reduces to its inputs by construction, so the score is moderate.

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

No new physical entities are introduced; the novel elements are trainable architectural components. The free parameters and axioms listed above are the hand-chosen or asserted elements the central claims depend on, beyond standard transformer machinery and the AWGN channel model.

free parameters (4)
  • lambda_B (BLER regularization weight) = 0.1
    Hand-selected weight for the non-differentiable BLER term in the KLD loss; affects the training objective but is not derived from first principles.
  • p_b (confidence threshold for experience reflow) = 0.55
    Hand-selected threshold that decides which misclassified samples are retained and relabeled with reference decoder outputs.
  • L_m (experience buffer size) = 128
    Hand-selected memory buffer size that limits how many boundary samples are reused across epochs.
  • Architecture dimensionality hyperparameters (d_m, N_max, d_f, N_b) = Ambiguous values in Table I
    Core architecture dimensions are chosen by hand and listed in Table I, but the column alignment is unclear, making the exact values hard to reproduce.
assumptions (4)
  • domain assumption Frozen bit positions are selected via Gaussian approximation of subchannel reliability.
    Sections II-B and VI-A rely on GA-based LLR analysis from [27]; the exact finite-length polarization order can differ from this approximation.
  • ad hoc to paper Front-padding preserves the polar code tree structure under an affine bit transformation.
    Section III-A asserts that Eq. (14) keeps padded zeros and valid codewords separated across the tree, which underpins cross-length generalization but is not proved in detail.
  • ad hoc to paper The energy-based model in Section IV-A justifies trainability and decoding capability of the LAT architecture.
    Equations (32) through (38) apply Tweedie's formula and a white-box transformer analysis, but the extension to latent attention with LPE-generated Q and K is asserted rather than derived.
  • domain assumption Reported performance is measured on a configuration that has received additional fine-tuning.
    Section VI-B adds 100 epochs of fine-tuning for N=16, k=8 before testing, and Section VI-C admits degradation when fine-tuning is omitted; the central claims presume this adaptation step.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent-attention Based Transformer for Near ML Polar Decoding in Short-code Regime." pith.science (2026). https://pith.science/paper/YTBANNJG

@misc{pith2026250714951,
  author       = {Pith},
  title        = {Pith review of: Latent-attention Based Transformer for Near ML Polar Decoding in Short-code Regime},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YTBANNJG}},
  note         = {Machine review of arXiv:2507.14951}
}
read the original abstract

Transformer architectures have emerged as promising deep learning (DL) tools for modeling complex sequence-to-sequence interactions in channel decoding. However, current transformer-based decoders for error correction codes (ECCs) demonstrate inferior performance and generalization capabilities compared to conventional algebraic decoders, especially in short-code regimes. In this work, we propose a novel latent-attention based transformer (LAT) decoder for polar codes that addresses the limitations on performance and generalization through three pivotal innovations. First, we develop a latent-attention mechanism that supersedes the conventional self-attention mechanism. This architectural modification enables independent learning of the Query and Key matrices for code-aware attention computation, decoupling them from the Value matrix to emphasize position-wise decoding interactions while reducing context correlation interference. Second, we devise an advanced training framework incorporating three synergistic components: entropy-aware importance sampling that emphasizes low-probability regions in the signal constellation space, experience reflow that introduces empirical labels to improve characterization of decoding boundaries, and dynamic label smoothing for likelihood-based regularization. Third, we propose a code-aware mask scheme which allows dynamic adaptation for varying code configurations. Numerical evaluations demonstrate that the proposed LAT decoder achieves near maximum-likelihood (ML) performance in terms of both bit error rate (BER) and block error rate (BLER) for short-length polar codes. Furthermore, the architecture exhibits robust generalization capabilities across diverse code rates and code lengths.

Figures

Figures reproduced from arXiv: 2507.14951 by the authors.

Figure 1
Figure 1. An end-to-end system model. The message bit sequence [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. The LAT decoder architecture. Each decoding layer incorporates an LPE that simultaneously generates a latent positional Query encoding [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. (a) Post-padding and (b) Front-padding for N = 4 and Nmax = 8 with the corresponding code trees, where the gray boxes represent padding zeros. Through front-padding, the absolute bit indices are preserved across varying code lengths, which is consistent with their inherent node positions in the padded code tree structures, where gray nodes represent the front-padding zeros and white nodes represent the valid codewor… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: The code-aware masking scheme for Nmax = 8, N = 4, and k = 2. Gray boxes are masked out whereas yellow boxes retain their values. Through this mask, the frozen bits and padding zeros are isolated from other input values while the attention weights over the message rows…
Figure 6
Figure 6. Figure 6: The masking pattern shift from codes with [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The masking pattern shift from codes with [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Performance of the LAT decoder, the ECCT decoder, the SC decoder, the SCL (Lsc = 4) decoder, and the ML decoder with code configuration of N = 16 and k = 8, (a) BER, (b) BLER. listed in Table I. The training dataset is obtained through entropy-aware importance sampling…
Figure 9
Figure 9. Figure 9: Performance of the LAT decoder and the SCL (Lsc = 4) decoder, (a) BER and (b) BLER across code configurations of {N = 16, k = 8}, {N = 16, k = 12}, {N = 8, k = 4}, and {N = 8, k = 6}; (c) BER and (d) BLER across code rate R ∈ [0.25, 0.75] under Eb/N0 = 5 dB for N ∈ {8,…
Figure 10
Figure 10. Figure 10: Performance of the LAT decoder, and the weakened LAT decoder with code configurations of N = 16 and k = 8, (a) BER, (b) BLER. D. Benefit of the Modified Training Strategy Since error labels are inevitably introduced into the training dataset through sampling over nois…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 30 canonical work pages

  1. [1]

    On the road to 6G: Visions, requirements, key tech- nologies and testbeds,

    C. Wang, et al. , “On the road to 6G: Visions, requirements, key tech- nologies and testbeds,” IEEE Commun. Surv. Tutor. , vol. 25, no. 2, pp. 905–974, Feb. 2023. 14

    work page 2023

  2. [2]

    Disentangled representation learning empowered CSI feedback using implicit channel reciprocity in FDD massive MIMO,

    W. Xu, J. Wu, S. Jin, X. You, and Z. Lu, “Disentangled representation learning empowered CSI feedback using implicit channel reciprocity in FDD massive MIMO,” IEEE Trans. Wirel. Commun., vol. 23, no. 10, pp. 15169–15184, Oct. 2024,

    work page 2024

  3. [3]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, pp. 379–423 and 623–656, July and Oct. 1948

    work page 1948

  4. [4]

    Channel coding: The road to channel capacity,

    D. J. Costello and G. D. Forney, “Channel coding: The road to channel capacity,” Proc. IEEE, vol. 95, no. 6, pp. 1150–1177, Jul. 2007

    work page 2007

  5. [5]

    Edge learning for B5G networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,

    W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y . C. Eldar, and M. Debbah, “Edge learning for B5G networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,” IEEE J. Sel. Topics Signal Process. , vol. 17, no. 1, pp. 9–39, Jan. 2023,

    work page 2023

  6. [6]

    Near shannon limit error- correcting coding and decoding: Turbo-codes,

    C. Berrou, A. Glavieux, and P. Thitimajshima, “Near shannon limit error- correcting coding and decoding: Turbo-codes,” in Proc. IEEE Int. Conf. Commun., Piscataway, New Jersey, May 1993, pp. 1064–1070

    work page 1993

  7. [7]

    Low-density parity-check codes,

    R. G. Gallager, “Low-density parity-check codes,” IRE Trans. Info. Theory, vol. 8, no. 1, pp. 21–28, Jan. 1962

    work page 1962

  8. [8]

    Channel polarization: A method for constructing capacity achieving codes for symmetric binary-input memoryless channels,

    E. Arıkan, “Channel polarization: A method for constructing capacity achieving codes for symmetric binary-input memoryless channels,” IEEE Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009

    work page 2009

Show all 35 references
  1. [9]

    On the rate of channel polarization,

    E. Arıkan and E. Telatar, “On the rate of channel polarization,” in Proc. IEEE Int. Symp. Inf. Theory , Seoul, Korea, Jul. 2009, pp. 1493–1495

    2009

  2. [10]

    CRC-aided decoding of polar codes,

    K. Niu and K. Chen, “CRC-aided decoding of polar codes,” IEEE Commun. Lett., vol. 16, no. 10, pp. 1668–1671, Oct. 2012

    2012

  3. [11]

    6G: A welcome chance to unify channel coding?

    M. Geiselhart, F. Krieg, J. Clausius, D. Tandler, and S. ten Brink, “6G: A welcome chance to unify channel coding?” IEEE BITS Inf. Theory Mag., vol. 3, no. 1, pp. 67–80, Oct. 2023

    2023

  4. [12]

    Channel coding for 6G extreme connectivity requirements, capabilities and fundamental tradeoffs,

    H. Zhang and W. Tong, “Channel coding for 6G extreme connectivity requirements, capabilities and fundamental tradeoffs,” IEEE BITS Inf. Theory Mag., vol. 3, no. 1, pp. 54–66, Oct. 2023

    2023

  5. [13]

    Recent advances in deep learning for channel coding: A survey,

    T. Matsumine and H. Ochiai, “Recent advances in deep learning for channel coding: A survey,” IEEE Open J. Commun. Soc. , vol. 5, pp. 6443–6481, Oct. 2024

    2024

  6. [14]

    Multilayer feedforward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural Netw., vol. 2, no. 5, pp. 359–366, Jul. 1989

    1989

  7. [15]

    In-datacenter performance analysis of a tensor processing unit,

    N. P. Jouppi, et al. , “In-datacenter performance analysis of a tensor processing unit,” in Proc. Int. Symp. Comput. Archit. , New York, NY , USA, Jun. 2017, pp. 1–12

    2017

  8. [16]

    Comparing energy efficiency of CPU, GPU and FPGA implementations for vision kernels,

    M. Qasaimeh, K. Denolf, J. Lo, K. Vissers, J. Zambreno, and P. H. Jones, “Comparing energy efficiency of CPU, GPU and FPGA implementations for vision kernels,” in Proc. IEEE Int. Conf. Embed. Softw. Syst. , Las Vegas, NV , USA, Jun. 2019, pp. 1–8

    2019

  9. [17]

    List decoding of polar codes,

    I. Tal and A. Vardy, “List decoding of polar codes,” Inf. Theory Proc. , vol. 61, no. 5, pp. 1–5, Mar. 2011

    2011

  10. [18]

    List successive cancellation decoding of polar codes,

    K. Chen, K. Niu, and J. R. Lin, “List successive cancellation decoding of polar codes,” Electron. Lett., vol. 48, no. 9, pp. 500–501, Apr. 2012

    2012

  11. [19]

    Stack decoding of polar codes,

    K. Niu and K. Chen, “Stack decoding of polar codes,” Electron. Lett., vol. 48, no. 12, pp. 695–696, Jun. 2012

    2012

  12. [20]

    On deep learning- based channel decoding,

    T. Gruber, S. Cammerer, J. Hoydis, and S. ten Brink, “On deep learning- based channel decoding,” in Proc. Conf. Inf. Sci. Syst. , Baltimore, MD, USA, Mar. 2017, pp. 1–6

    2017

  13. [21]

    Performance evaluation of channel decoding with deep neural networks,

    W. Lyu, Z. Zhang, C. Jiao, K. Qin, and H. Zhang, “Performance evaluation of channel decoding with deep neural networks,” inProc. IEEE Int. Conf. Commun. , Kansas City, MO, USA, May 2018, pp. 1–6

    2018

  14. [22]

    Performance analysis of deep learning based on recurrent neural networks for channel coding,

    R. Sattiraju, A. Weinand, and H. D. Schotten, “Performance analysis of deep learning based on recurrent neural networks for channel coding,” in Proc. IEEE Int. Conf. Adv. Netw. Telecommun. Syst. , Indore, India, Dec. 2018, pp. 1–6

    2018

  15. [23]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. Adv. Neural Inf. Process. Syst. , Long Beach, California, USA, Dec. 2017, pp. 6000–6010

    2017

  16. [24]

    Error correction code transformer,

    Y . Choukroun and L. Wolf, “Error correction code transformer,” in Proc. Conf. Neural Inf. Process. Syst. , New Orleans, LA, USA, Dec. 2022, pp. 38695–38705

    2022

  17. [25]

    CrossMPT: Cross-attention message-passing transformer for error correcting codes,

    S.-J. Park, H.-Y . Kwak, S.-H. Kim, Y . Kim, and J.-S. No, “CrossMPT: Cross-attention message-passing transformer for error correcting codes,” arXiv preprint arXiv:2405.01033 , 2024

    2024 arXiv

  18. [26]

    On the design and performance of machine learning based error correcting decoders,

    Y . Yuan, P. Scheepers, L. Tasiou, Y . G ¨ultekin, F. Corradi, and A. Alvarado, “On the design and performance of machine learning based error correcting decoders,” arXiv preprint arXiv:2410.15899 , 2024

    2024

  19. [27]

    Efficient design and decoding of polar codes,

    P. Trifonov, “Efficient design and decoding of polar codes,” IEEE Trans. Commun., vol. 60, no. 11, pp. 3221–3227, Nov. 2012

    2012

  20. [28]

    Beyond dis- crete selection: continuous embedding space optimization for generative feature selection,

    M. Xiao, D. Wang, M. Wu, P. Wang, Y . Zhou, and Y . Fu, “Beyond dis- crete selection: continuous embedding space optimization for generative feature selection,” in Proc. IEEE Int. Conf. Data Min. (ICDM) , Shanghai, China, Dec. 2023, pp. 688–697

    2023

  21. [29]

    Mish: A self regularized non-monotonic activation function,

    D. Misra, “Mish: A self regularized non-monotonic activation function,” arXiv preprint arXiv:1908.08681 , 2019

    1908 arXiv

  22. [30]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas, NV , USA, Jun. 2016, pp. 770–778

    2016

  23. [31]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  24. [32]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas, NV , USA, Jun. 2016, pp. 2818–2826

    2016

  25. [33]

    Tweedie’s formula and selection bias,

    B. Efron, “Tweedie’s formula and selection bias,” J. Am. Stat. Assoc. , vol. 106, no. 496, pp. 1602–1614, Dec. 2011

    2011

  26. [34]

    A fast iterative shrinkage-thresholding algorithm with application to wavelet-based image deblurring,

    A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm with application to wavelet-based image deblurring,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. , Taipei, Taiwan, China, Apr. 2009, pp. 693–696

    2009

  27. [35]

    White-box transformers via sparse rate reduction,

    Y . Yu, S. Buchanan, D. Pai, T. Chu, Z. Wu, S. Tong, B. D. Haeffele, and Y . Ma, “White-box transformers via sparse rate reduction,” in Proc. Adv. Neural Inf. Process. Syst., New Orleans, Louisiana, USA, Dec. 2023, pp. 9422–9457

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.