Boosting Learning for LDPC Codes to Improve the Error-Floor Performance

Low-density parity-check (LDPC) codes have been successfully commercialized in communication systems due to their strong error correction capabilities and simple decoding process. However, the error-floor phenomenon of LDPC codes, in which the error rate stops decreasing rapidly at a certain level, presents challenges for achieving extremely low error rates and deploying LDPC codes in scenarios demanding ultra-high reliability. In this work, we propose training methods for neural min-sum (NMS) decoders to eliminate the error-floor effect. First, by leveraging the boosting learning technique of ensemble networks, we divide the decoding network into two neural decoders and train the post decoder to be specialized for uncorrected words that the first decoder fails to correct. Secondly, to address the vanishing gradient issue in training, we introduce a block-wise training schedule that locally trains a block of weights while retraining the preceding block. Lastly, we show that assigning different weights to unsatisfied check nodes effectively lowers the error-floor with a minimal number of weights. By applying these training methods to standard LDPC codes, we achieve the best error-floor performance compared to other decoding methods. The proposed NMS decoder, optimized solely through novel training methods without additional modules, can be integrated into existing LDPC decoders without incurring extra hardware costs. The source code is available at https://github.com/ghy1228/LDPC_Error_Floor .

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer UbgL7/10 · confidence 5/52023-06-17

Summary

1. In the proposed work authors proposed a. Neural Min-Sum (NMS) decoders b. NMS decoder with block-wise training schedule that locally trains a block of weights while retraining the preceding block. c. The different weights are assigned to the unsatisfied check nodes and training is carried out. 2. The contribution of the proposed work are. a. Boosting learning using uncorrected codewords b. Block-wise training schedule with retraining c. Weight sharing technique with dynamic weight allocation 3. Sufficient results are presented and discussed.

Strengths

Novelty of the work is good. Sufficient results are presented and discussed.

Weaknesses

The experimetnal set up should be elaboprated more with parameters selected for training.

Questions

1. How the weights of the NMS decoders are initialized. 2. Is it possible to design ASIC hardware of NMS decoder 3. What about the latency of the decoder. 4. The neural decoders with recent state of art techniques with their advantages should be discussed in detail. 5. The experimental set up should be clearly mentioned with parameters used for the training. 6. How exactly the weights are updated in the training process of LDPC CODES mentioned in the figure 1 and what is the impact of weights in the training process should be clearly mentioned.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Since the authors mentioned that proposed work can be taken into hardware architecture. But clear architectute of the proposed method is not highlighted.

Reviewer jBEt6/10 · confidence 4/52023-06-21

Summary

The paper proposes a training framework for the NMS decoder of LDPC codes in order to enhance the error-floor performance of these codes. \ The NMS decoder iterations are split into two cascaded parts, the first so-called based decoder is trained for decoding waterfall parts, and the second (post decoder) is for decoding uncorrected codewords caused by the error-floor.\ The paper suggests the use of a training schedule reusing learned weights to tackle the vanishing gradient of many iterations NMS.\ The proposed decoder using the trained weights showed improved error-floor performances for several standard LDPC codes and can be efficiently integrated into current NMS solutions.

Strengths

The paper addresses an important issue present in BP/MS-based decoders on these codes. The solution is simple and does not require any architectural modification. The performance seems important and is obtained without adding any complexity compared to the original NMS decoder.

Weaknesses

#### NMS I am not sure the comparison with NMS is fair for multiple reasons. 1) The training of NMS can be performed in several ways that should not allow zero loss. For example, the observations in lines 224 and 225 can be easily solved by using large batches and/or using batches spanning many Eb/N0 values. 2) The vanishing gradient phenomenon has been solved in [5] using a multi-loss objective, which clearly prevents the gradients to collapse at each layer. Also, in Figure 4b) NMS and the method should be compared at equal number of iterations. NMS's test FERs with l=40 and l=50 are not that far from the l=50 proposed method FER. 3) NMS remains non-negligibly better in the waterfall region. 4) Many NMS algorithms have emerged since [5] and I am wondering if the comparison with the original NMS [5](2018) is fair and conclusive in order to judge the paper. #### Clarity: Figure 2 is not clear. For example, the differentiation between the NMS and the neural network is not clear; the base and post NN should be linked by an arrow. The paper contains a very large amount of text which can become cumbersome. I believe adding clear Algorithm(s) would simplify the comprehension a lot. #### Misc: Typos: line 88: we a new \ line 121: protogrph

Questions

1) Please address the NMS subsection from the Weaknesses section.\ It would be beneficial to have an ablation study of the final accuracy for NMS augmented with the different components of the proposed methods. Also, different training strategies of the NMS must be considered. 2) Please address the clarity and Misc from the Weaknesses section. I believe adding clarity, addressing the ablations, and adding more experiments would increase the quality, impact, and rating of the paper.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

No limitations are discussed.

Reviewer MFD66/10 · confidence 4/52023-07-04

Summary

The author studies LDPC code's neural min-sum decoder's error floor problem, by solely change the training method: (1) boosting learning method, which use first network to do majority of decoding to achieve waterfall region, and second network deal with small error residuals. (2) relief deep decoding iteration's vanish gradient problem by block-wise training. (3) bundle weight in dynamic manner, to reduce the number of weights to be trained. With above techniques, the experiment runs on varies LDPC settings, shows that proposed NMS has better performance on error-floor regions, with same computational resource as NMS. Since NMS and MS have similar complexity, it is believed that proposed NMS could be possibly adapted to future LDPC decoder implementation.

Strengths

(1) The paper works directly on training of NMS algorithm, which requires minimal change on existing low complexity NMS solutions. All three training methods are reasonable and are supported by experiments. The residual learning (termed as boosting learning) empirically shows advantage on error-floor regions. The block-wise training schedule does improve performance with large number of iterations. Finally the weight sharing technique adds a structural sparse limitation to neural decoder, which only uses a small portion of weights. (2) The evaluation is extensive, on multiple LDPC settings at Figure 6, we see that proposed NMS method shows better error-floor performance, compared to existing NMS and canonical decoders.

Weaknesses

Here are my concerns: (1) The major contribution on Machine Learning sense, is limited. The proposed boosting learning method is common, as residual learning in ML sense; the proposed training schedule and weight sharing are also widely-used ML methods. The core contribution is on applying ML techniques in the context of channel coding field. In some sense, the targeted audience shall be channel coding researchers, rather than general ML researchers. (2) The experiments are not conducted in long block length (length>1000, which is common for QC-LDPC), which is widely used in 5G/WIFI systems. It might be interesting to check the performance on long block length, where block length coding gain is significant, and NMS shows less advantage against canonical methods.

Questions

N/A

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer zXhE7/10 · confidence 5/52023-07-06

Summary

The paper proposes training methods to optimize neural min-sum (NMS) decoders that are robust to the error-floor phenomenon of LDPC codes. The proposed methods include: (1) dividing the decoding network into two neural networks and training the post network to be specialized for uncorrected codewords that failed in the first network; (2) introducing a block-wise training schedule that locally trains a block of weights while retraining the preceding block; and (3) assigning different weights to unsatisfied check nodes. The proposed methods are applied to standard LDPC codes, and the results show that they achieve the best decoding performance in the error-floor region compared to other decoding methods in the literature. The proposed NMS decoder is designed only by modifying the training methodology, without adding any additional modules. Therefore, it can be seamlessly incorporated into the well-established architectures of LDPC decoders and can be immediately utilized in practical domains that demand exceptionally low error rates.

Strengths

The paper introduce novel ideas: 1. Boosting learning using uncorrected codewords - The authors propose a boosting learning technique for LDPC code decoding. The technique divides the decoding network into two cascaded networks, where the first network focuses on the waterfall performance and the second network focuses on the error-floor region. The second network is able to correct uncorrected codewords that are not corrected by the first network due to the error-floor phenomenon. This results in a significant performance improvement in the error-floor region. 2. Block-wise training schedule with retraining - The authors propose a new training schedule for NMS decoders that mitigates the vanishing gradient problem. The proposed schedule divides the entire decoding iterations into sub-blocks and trains the blocks in a sequential manner. Additionally, the weights trained from previous blocks are retrained to escape from local minima. This results in a significant performance improvement over the one-shot training method and the iter-by-iter schedule. 3. Weight sharing technique with dynamic weight allocation - The authors propose a new weight sharing technique for NMS decoders that improves the performance in the error-floor region. The technique dynamically assigns different weights to unsatisfied check nodes (UCNs) and satisfied check nodes (SCNs) in the decoding process. This results in a significant performance improvement with a 2.6% reduction in the number of weights to be trained.

Weaknesses

.

Questions

1. Does the proposed method can be also beneficial to improve the results of neural decoder for other types of codes? such as BCH and Polar? 2. How does the model behave with other type of noise (other than Gaussian)? 3. Can you tie the weights of the different network to get improvement in terms of model size? 4. What happen if you take more then two NN? does the performance improve?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

.

Reviewer uaWt5/10 · confidence 4/52023-07-11

Summary

This paper presents novel training techniques for the NMS decoder of LDPC codes aimed at enhancing performance in the error floor region. The proposed decoding methods comprise two stages: a base decoder and a post decoder designed specifically for uncorrected codewords from the base decoder. In order to tackle the issue of vanishing gradient during training, a block-wise training schedule is introduced in this paper. By assigning distinct weights to unsatisfied check nodes, the error floor can be lowered with a small number of weights to be trained.

Strengths

1. The impact of the proposed decoding technique is substantial, as evidenced by the figures which clearly demonstrate a significant improvement in code performance, particularly in the error floor region. 2. The analysis and interpretation of the results in Figure 3 were particularly valuable, providing a clear understanding of the distinctions between the methods employed for selecting training samples.

Weaknesses

1. The proposed decoding method in this paper builds upon the concept of two-stage decoding and incorporates a post decoder that complements the base decoder. It is important to acknowledge that this idea is not entirely new within the field of coding theory for LDPC codes, as there has been prior research exploring similar approaches. The followings are some of the examples sharing a similar philosophy and it would be valuable to mention them as relevant previous work in this direction. S. Yang, Y. Han, X. Wu, R. Wood and R. Galbraith, "A soft decodable concatenated LDPC code," 2015 IEEE International Magnetics Conference (INTERMAG), Beijing, China, 2015 J. Oh, J. Ha, H. Park and J. Moon, "RS-LDPC Concatenated Coding for the Modern Tape Storage Channel," in IEEE Transactions on Communications, vol. 64, no. 1, pp. 59-69, Jan. 2016 H. Park and J. Moon, "Improving SSD Read Latency via Coding," in IEEE Transactions on Computers, vol. 69, no. 12, pp. 1809-1822, 1 Dec. 2020 2. The clarity of the presentation style could be enhanced. For example, the specific problem addressed by the paper, such as the type of channel considered, is not explicitly stated. The reader must infer this information primarily from the related work section. Providing a clear and explicit description of the problem, including the type of channel under consideration, would greatly improve the overall clarity of the paper. 3. Nitpicks - It would greatly enhance the presentation if the theoretical performance limit based on the code rate could be included in the plots, such as in Figures 3 and 6. - Typo: On line 121, page 3, "protogrph" should be corrected to "protograph".

Questions

In order to harness the potential of boosting learning, this study introduces a division of the decoding network into two components, with the post decoding stage aiding the base decoding process. A natural extension of this approach is to explore the possibility of incorporating more than two decoders. It would be intriguing to learn from the authors whether they have considered this scenario and if they can provide insights into the expected outcomes. Including discussions and potential answers to this question would enhance the manuscript's appeal and stimulate further interest in the research.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

N/A - This work does not appear to have any discernible negative societal impact.

Reviewer jBEt2023-08-10

Thank you for your very clear and thorough answer. 1) Thank you for the experiment. What is the impact of multi-loss applied with your method? Do you assume it won't help? 2) Comparison with other works: the comparison is not very pertinent since much better neural decoders than [15] have been recently developed (e.g., [A, B] are related to the authors of [5] and [15]). However, I admit their capacities may not be straightforwardly equivalent (e.g., [C] is not performing as well but is a more lightweight model). 3) It seems the careful selection of the data solely seems to have an impressive and surprising impact on the performance. - Do you know what is the real impact of training and fine-tuning on the whole model (base+post)? (maybe with different step-sizes) - Do you know if other data-selection policies during training have already been investigated? (related works focus on weights repartitions). Thank you [A] *Autoregressive Belief Propagation for Decoding Block Codes* [B] *Error Correction Code Transformer* [C] *Graph Neural Networks for Channel Decoding*

Authorsrebuttal2023-08-13

Reply to the comments by jBEt

1. In the error floor scenario with FER metric, the proposed block-wise method shows the best results, but the multi-loss method is the second most effective. The multi-loss approach has the advantage of enabling learning over a large number of iterations without suffering from the vanishing gradient problem unlike the at-once method. Also, in contrast to the iter-by-iter method, it incorporates multiple iterations in the loss function, which more effectively eliminates error patterns in the error floor. As a result, the multi-loss method achieves a lower FER compared to the at-once and iter-by-iter methods. We will mention these advantages in the revised version. 2. Thank you for the valuable comment. We will add a comparison with other works [A], [B]. (A comparison with [C] is not conducted as the source code is not publicly accessible and an apple-to-apple comparison is little hard) As is known, in the waterfall region, the performance of [B] demonstrates state-of-the-art results. Especially for high-density codes such as BCH and polar codes, there's a significant performance gap when compared with BP and MS research series (Hyper [15], AR [A], Boosting [Proposed]), whereas the gap is relatively minimal for LDPC codes. However, [B] uses an entirely different architecture based on transformer, which is not based on BP and MS algorithms, making its practicality somewhat limited as of now. Additionally, due to its intricate network structure, it's more complex than NMS decoders in terms of training time, memory requirement, and decoding complexity. In practice, the ECCT has difficulty training for the target codes of this paper, particularly for codes of several hundred lengths with very low error rates. In our GPU environment, training for the WiMax code (N=576, K=432) is feasible using only `shallow` ECCT (N=2, d=32), and at this depth, there's no performance improvement over the BP & MS series of works. On the other hand, our proposed boosting learning is trainable even in very deep networks, around 100 iterations, and long codes 1000 in length. Although our focus in this work was on improving the error floor performance of neural min-sum decoders, our approach could potentially be applied to other decoding architectures (e.g., ECCT). Application of our boosting learning to ECCT would be an enticing follow-up as a future study. We conjecture that the boosting learning via smart selection and arrangement of training samples can be effective on a large class of architectures. In summary, the contributions of our work are as follows: i) Primary contribution: Achieving SOTA in terms of error floor performance in practical situations like the NMS decoder, standard LDPC codes. ii) Secondary contribution: Achieving a competitive waterfall performance for short codes by modifying only the training method. (will be discussed in the supplementary file) iii) Additional benefit: A flexible methodology applicable to different channels and decoder architectures. 3. As you mentioned, careful selection of data samples can lead to a significant improvement in decoding performance through the proposed boosting approach. The ablation study underscores this substantial performance boost (as seen when comparing the first and second rows of the ablation study: from $3.14 \times 10^{-6}$ to $3.31 \times 10^{-7}$. Most research has focused on proposing new decoder architectures or decoding algorithms to improve decoding performance. In contrast, our proposed boosting approach demonstrates that the decoding performance can be significantly enhanced by training carefully selected data samples. We believe that the proposed boosting approach offers a promising new research direction that can improve decoding performance while minimizing the additional decoding complexity. I've explored fine-tuning of some hyper-parameters (learning rate, initial values, batch size) but didn't observe significant effects. Instead, we will include experimental results showing valid effects, such as the usage of VN weights and splitting of the number of iterations between base and post. Research on data selection has been conducted in “active learning” [16] and “BP-RNN diversity” [9]. [16] employs a single-stage decoding and gathers data samples with a moderate number of errors, inspired by active learning. However, [16] significantly differs from our proposed boosting learning method, which uses a two-stage decoding process and collects the failed codeword samples during the first stage. BP-RNN Diversity [9] requires enumerating trapping sets beforehand and because of this, it can only be applied to short codes. In contrast, the sampling of uncorrected words in the proposed boosting learning has an O(n) complexity, making it feasible for longer codes. More importantly, they focused on the waterfall performance. We will clarify this point in the revised manuscript. Thank you again for the valuable comments.

Reviewer jBEt2023-08-16

Thank you for your answer. I believe the rebuttal improved the paper and even if the neural contribution is not substantial, the problem and the simple solution may be impactful enough.

Reviewer UbgL2023-08-13

I thank the authors for the detailed response. I think the majority of my concerns have been addressed.

Reviewer MFD62023-08-18

Keep my score the same

After reading the response from the author, I decide to keep my rating.

Area Chair RPzh2023-08-18

Thank you for the rebuttal

Dear authors, thank you for providing a rebuttal. Some of the reviewers have already replied, so this is just to let you know that I am in contact with the remaining ones as well. Best, Your AC

Reviewer zXhE2023-08-18

Thank you addressing my concerns. The answers are satisfactory.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC