Neural Window Decoding of Spatially Coupled LDPC Codes: Exploiting Window Structure for Learning and Adaptation
Authors: Dae-Young Yun, Hee-Youl Kwak, Sang-Hyo Kim, Yongjune Kim, Jong-Seon No
Organizations: Department of Electrical, Electronic and Computer Engineering, University of Ulsan, Ulsan 44610, South Korea · Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon 16419, South Korea · Department of Electrical Engineering at Pohang University of Science and Technology (POSTECH), Pohang, Gyeongbuk 37673, South Korea · Department of Electrical and Computer Engineering, INMC, Seoul National University, Seoul 08826, South Korea
We propose a neural window decoder (NWD) for spatially coupled low-density parity-check (SC-LDPC) codes. The proposed methods adapt neural min-sum decoding to the structure and sequential operation of window decoding. First, target-oriented training restricts the loss to the variable nodes whose decisions are required at each window stage. Second, learned damping factors identify check-node updates that can reuse messages from the preceding iteration. Finally, adaptive weight switching selects weights trained for unreliable inherited information when a previous-window failure is detected. For a length-20,000(3,6)-regular SC-LDPC code, the target-oriented NWD achieves performance comparable to sum-product window decoding while using min-sum updates. On the reduced target graph, scheduling based on regularized damping factors yields an estimated computational cost 48% lower than that of full-graph weighted min-sum decoding, while maintaining comparable decoding performance. Adaptive weight switching improves decoding performance by 35% at 3.0 dB relative to plain weights. These learning methods exploit the properties of window decoding to improve the reliability and computational efficiency of SC-LDPC convolutional codes.
Figures & tables
Fig. 1: (a) Protograph construction of the terminated (3,6) -regular SC-LDPC code with w=2 and L=4 by edge spreading. (b) Window decoding with W=6 and one target position. The dotted box denotes the window at the third stage; the dark-filled VNs are the target VNs, and the arrows denote inherited LLRs from earlier decoded blocks.
Fig. 2: The terminated first window with W=3 and z=1 , unrolled for ℓ=2 iterations. Solid connections are retained for the target outputs; light-gray nodes and dashed connections can be omitted. The labels ei identify Tanner-graph edges and their corresponding directed messages. Repeated input circles denote the same channel LLRs used to compute the VN updates and output LLRs.
Fig. 3: Retained-weight ratio of NWD relative to NBD for the (3,6) -regular, w=2 , T=1 graph. Each cell shows one combination of iteration budget ℓ and window width W ; darker blue denotes a smaller retained fraction and hence greater reduction. Step boundaries indicate the 50%, 75%, and 90% retained-weight thresholds. The outlined cell marks the main setting W=10 , ℓ=10 .
Fig. 4: Error-rate performance for w=2 , z=100 , W=10 , T=1 , and ℓ=10 : (a) single-window BLER and (b) full-chain FER for L=100 ( Ln=20,000 ). WD, NBD, and NWD use MS-based updates; only WD BP uses sum-product BP updates. WD uses one fixed shared weight, NBD is trained on all window outputs, and NWD is trained on the target outputs and retains the target subgraph.
Fig. 5: CN-update ranking metrics for W=10 , T=1 , and ℓ=10 : (a) NWD weights, (b) damping factors learned with λ=0 , and (c) regularized damping factors with λ=1 . Panels (b) and (c) train only damping factors while fixing the weights in (a). Numbers indicate the first 29 updates in each rest order. Gray cells lie outside the target subgraph.
Fig. 6: Required Eb/N0 at FER =10−1 versus rested updates R for three scheduling methods. The upper rows show computational costs; dotted and dashed lines denote Triangle ( R=29 ) and WD ( R=0 ), respectively.
Fig. 7: Weight sets for adaptive weight switching: (a) plain weights wC(ℓ) , (b) barrier weights wC(ℓ) , and (c) their difference wC(ℓ)−wC(ℓ) .
Fig. 8: Adaptive weight switching for the next window. The UCN indicator from the previous window chooses the plain weights for fs=0 and barrier weights for fs=1 . The first window uses the plain weights.
Fig. 9: Adaptive weight switching for the L=100 , z=100 , W=10 , T=1 , ℓ=10 code: (a) position-averaged q1 and (b) average BLER for Plain only, UCN-Switch, and Genie-Switch.
Complexity per protograph
Equivalent cost
Stored weights
Addition
Comparison
Sign mul.
Real mul.
LUT
(A+C+0.3S
A
C
S
P
T
+4P+6T)z
WD BP
1080
0
0
1620
1080
1404000 (+236%)
0
WD
1080
610
1080
540
0
417400 (reference)
1
NBD
1080
610
1080
540
0
417400 (reference)
100
NWD
888
498
888
444
0
342840 (-18%)
84
TABLE I: Operation and stored-weight counts for a left-terminated reference window with (WN,WM,z,ℓˉ)=(20,10,100,10) . Per-protograph operation counts are summed before lifting; the equivalent computational cost includes z .
Fig. 10: Complexity–reliability tradeoffs: left, FER at 2.4 dB; right, average BLER at 3.0 dB. Computational costs follow the model in Table I and are normalized to WD. Scheduling uses R=15 .
Dae-Young Yun (Member, IEEE) is with the Department of Electrical, Electronic and Computer Engineering, University of Ulsan, Ulsan, South Korea. His research interests include channel coding, iterative decoding, and learning-aided communication systems.
Table 12
Hee-Youl Kwak (Member, IEEE) is with the Department of Electrical, Electronic and Computer Engineering, University of Ulsan, Ulsan, South Korea. His research interests include error-control coding, iterative decoding, and machine learning for communication systems.
Table 13
Sang-Hyo Kim (Member, IEEE) is with the Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, South Korea. His research interests include channel coding, information theory, and communication systems.
Table 14
Yongjune Kim (Member, IEEE) is with the Department of Electrical Engineering, Pohang University of Science and Technology, Pohang, South Korea. His research interests include coding theory, iterative decoding, and reliable communication systems.
Table 15
Jong-Seon No (Life Fellow, IEEE) is with the Department of Electrical and Computer Engineering and the Institute of New Media and Communications, Seoul National University, Seoul, South Korea. His research interests include coding theory, cryptography, sequence design, and wireless communications.
Forward error correction is essential for reliable communication over noisy channels. Attention-based model-free neural decoders have shown strong performance for short codes, but their scalability to longer codes is limited by the quadratic memory and computational cost of attention. In this paper, we introduce the Mamba message-passing decoder (MMPD), an attention-free syndrome-based neural decoder for binary linear codes. MMPD retains the Tanner-graph structure of a message-passing decoder by performing local pairwise aggregation along variable-check edges. To enable efficient long-range information propagation, these local updates are combined with bidirectional Mamba state-space blocks. By avoiding dense attention matrices, MMPD scales more favorably for long codes in both memory and computation. Experiments on the (1056, 880) LDPC code show that MMPD achieves a 0.45 dB gain over the state-of-the-art CrossMPT decoder at a specified target bit error rate, while reducing memory consumption by a factor of 1.5. This reduction factor increases substantially for longer codes, demonstrating the applicability of MMPD to scalable neural decoding of practical long codes.
Quantum error correction is essential for fault-tolerant quantum computing. High-rate quantum low-density parity-check (qLDPC) codes have emerged as a promising approach to reducing the substantial physical qubit overhead of quantum error correction, enabling encoding rates exceeding 1/2 in blocks of approximately one thousand physical data qubits. However, their practical implementation requires decoders that combine high accuracy with low latency to keep pace with quantum hardware. Existing decoders of high-rate qLDPC codes can achieve high accuracy but often incur substantial computational latency, whereas neural network decoders have yet to be demonstrated for high-rate qLDPC codes. Here, we introduce neural triage decoding, a framework that uses syndrome consistency to adaptively allocate neural computation. A local neural message-passing network rapidly predicts error corrections, additional iterations reduce residual syndrome violations, and a rarely invoked diversification stage searches for consistent corrections to the remaining shots. We demonstrate that the Triage decoder can decode ultra-high-rate qLDPC codes encoding a few hundred logical qubits, including lifted-product and Kasai codes, achieving state-of-the-art accuracy comparable to the near optimal algorithmic decoders while reducing decoding latency by approximately one to two orders of magnitude. Our framework is inspired by a rigorous analysis of the exact bitwise marginal decoder, revealing a distance-independent syndrome-inconsistent error floor. In addition, we show the robustness of neural network decoders to mismatched error priors. Our results establish adaptive neural decoding as a promising route towards fast, accurate, and scalable quantum error correction in the near term.
Hong-Ye Hu, Zhenlai Wang, Binzhao Luo +2
Institute of Physics, Chinese Academy of Sciences, Beijing 100910, China · School of Physical Sciences, University of Chinese Academy of Sciences, Beijing 100049, China · School of Mathematical Sciences, Eastern Institute of Technology, Ningbo, Zhejiang 315200, China
Error correcting codes are a fundamental component of communication systems, enabling reliable data transmission over noisy channels. Recently, neural decoders have demonstrated notable performance gains over classical iterative algorithms. In this paper, we propose a Generalized Belief Propagation Network (GBPNet) that preserves message-passing symmetry while enhancing the decoder's learning capability. The architecture uses cascaded variable-node updates and learned element-wise message scaling while retaining the Tanner graph structure. We also introduce two training strategies, progressive training and high-SNR training, that improve decoding accuracy. Experimental results on several linear codes show that GBPNet achieves lower bit error rates than conventional belief propagation and the evaluated state-of-the-art neural decoders.