Technical Summary: Score Based Error Correcting Code Decoder (SB-ECC)

Problem Statement

Reliable digital communication relies on Error-Correcting Codes (ECC) to reconstruct transmitted codewords from noisy channel observations. While Maximum-Likelihood (ML) decoding is theoretically optimal, it is NP-hard for general linear block codes, making practical soft decoding a significant challenge.

Existing deep learning approaches for decoding generally fall into two categories, both with limitations:

1. Model-based decoders: These "unroll" iterative algorithms like Belief Propagation (BP) into trainable networks. While they benefit from strong inductive bias, their expressivity and scalability are often limited by the underlying iterative structure.
2. Model-free decoders: These learn the decoding function directly from data. However, they typically rely on preprocessing that maps the raw channel observation $y$ to its magnitude $|y|$ (reliability) combined with parity information (e.g., hard syndromes). This preprocessing folds the observation space, discarding directional information (the sign of $y$) which may be crucial for learning a continuous guidance field in $\mathbb{R}^n$. Furthermore, many model-free approaches struggle with the combinatorial explosion of the output space for longer block lengths, often leading to overfitting.

Additionally, diffusion-based decoders (e.g., DDECC) have emerged but often still rely on magnitude-based preprocessing and require explicit time or Signal-to-Noise Ratio (SNR) conditioning, which is often unavailable at inference time.

Methodology: SB-ECC

The authors propose SB-ECC (Score-Based Error-Correcting Code Decoder), a framework that casts soft decoding as a continuous-time denoising process guided by parity constraints.

Core Formulation

• Raw Signed Input: Unlike prior model-free decoders that use $|y|$, SB-ECC operates directly on the raw signed channel observation $y \in \mathbb{R}^n$. The authors argue that mapping $y \to |y|$ is non-invertible and collapses equivalence classes, potentially hindering the learning of a continuous score field.
• Variance-Exploding (VE) Construction: The channel transmission is modeled as a forward Variance-Exploding (VE) diffusion process where a clean BPSK codeword $x_0 \in \{\pm 1\}^n$ is perturbed by Gaussian noise: $y = x_0 + \sigma_{ch}\epsilon$.
• Score-Based Decoding: Decoding is formulated as reversing this diffusion process. The model learns a neural score function (or equivalently, an additive noise predictor $\hat{\epsilon}_\theta$) that estimates the noise vector given the noisy observation and code constraints.
• Parity Guidance: The decoder incorporates code structure via the parity-check matrix $H$. At each step of the denoising trajectory, a hard decision is made on the current state, and the syndrome $s = H \cdot \text{bin}(\text{sign}(x)) \pmod 2$ is computed. This syndrome is fed into the network to guide the denoising direction.

Architecture and Training

• Backbone: The model utilizes the CrossMPT architecture, which treats variable-node (VN) and check-node (CN) features as distinct modalities coupled via masked cross-attention blocks that respect the Tanner graph structure.
• Training Objective: The network is trained using Denoising Score Matching (DSM). It predicts the injected noise $\epsilon$ given a noisy sample $y = x_0 + \sigma(t)\epsilon$ and the corresponding syndrome. The loss is the mean squared error between the predicted and true noise.
• Time-Unconditional Training: A key practical innovation is that the model is trained across a range of noise levels without conditioning on the specific time step $t$ or SNR. This allows the model to operate without explicit SNR estimation at inference.

Inference

Decoding is performed by integrating the Probability-Flow Ordinary Differential Equation (PF-ODE) in continuous time, discretized in the noise scale $\sigma$-space.

• Process: Starting from the received vector $y$ (at an unknown noise level), the solver iteratively updates the state by subtracting the predicted noise scaled by the step size $\Delta\sigma$.
• Early Stopping: After each update, a hard decision is made. If the resulting codeword satisfies the parity checks (syndrome is zero), the process terminates early.
• Solver Flexibility: The framework supports different ODE solvers. The authors demonstrate that switching from the standard Euler solver to DPM-Solver (a higher-order solver) reduces the number of function evaluations required, thereby lowering latency while maintaining accuracy.

Key Contributions

1. SB-ECC Framework: Introduction of a time-unconditional, score-based decoder for linear block codes that learns a continuous additive-noise field directly on signed channel observations, guided by parity constraints.
2. Solver Optimization: Demonstration that replacing the Euler solver with DPM-Solver preserves error-rate performance ($-\ln(\text{BER})$) while significantly reducing end-to-end decoding time.
3. Empirical Superiority: The method achieves state-of-the-art performance across diverse code families (BCH, Polar, LDPC, MacKay, CCSDS) and block lengths without requiring SNR estimation.

Experimental Results

The authors evaluated SB-ECC across 42 code/SNR settings (various code types and lengths at $E_b/N_0 \in \{4, 5, 6\}$ dB) against strong baselines including BP, AR-BP, CrossMPT, and DDECC.

• Performance: SB-ECC achieved the best Bit Error Rate (BER) in 39 out of 42 entries.
• SNR Gain: Compared to the strongest competing baseline, SB-ECC showed an average SNR gain of 0.17 dB and a maximum gain of 0.46 dB.
• Latency-Accuracy Trade-off: By swapping the solver from Euler to DPM, the authors reduced end-to-end decoding time by an average of 8.86% (up to 12.82%) while maintaining the same $-\ln(\text{BER})$.
• Ablation Studies:
  • Time Conditioning: Removing explicit time/SNR conditioning did not hurt performance; in fact, the time-unconditional model performed best, suggesting the decoder can infer the noise regime directly from the signed observation.
  • Signed vs. Magnitude: Using magnitude-only inputs ($|y|$) caused the model to fail, with the network collapsing to a near-zero noise estimator and offering no coding gain over uncoded hard decisions. This validates the necessity of preserving directional information in the input.

Significance and Claims

The paper claims that SB-ECC demonstrates that learning decoding directly from raw signed observations is feasible and effective for mid-length codes, overcoming the sparsity issues often associated with model-free approaches. By framing decoding as continuous-time denoising, the method provides a flexible knob for trading off latency and accuracy at inference time without retraining. The work suggests that recent attention-based architectures, when combined with score-based generative modeling principles, can successfully learn continuous guidance fields for error correction, potentially offering a new paradigm for soft decoding that is both robust and computationally flexible.