Abstract
Hypernetworks were recently shown to improve the performance of message passing algorithms for decoding error correcting codes. In this work, we demonstrate how hypernetworks can be applied to decode polar codes by employing a new formalization of the polar belief propagation decoding scheme. We demonstrate that our method improves the previous results of neural polar decoders and achieves, for large SNRs, the same bit-error-rate performances as the successive list cancellation method, which is known to be better than any belief propagation decoders and very close to the maximum likelihood decoder.
Introduction
Polar codes are the first class of channel codes to provably achieve Shannon capacity and are a cornerstone of modern 5G standards. However, efficient decoding remains a significant challenge. Current approaches typically face a trade-off between accuracy and latency: Successive List Cancellation (SLC) offers high accuracy but is inherently serial, while Belief Propagation (BP) is highly parallelizable but generally less accurate.
Recent efforts to bridge this gap have utilized "Neural BP," replacing standard equations with learnable, yet static, weights. This work introduces a novel Gated Hypernet Decoder that addresses the lack of adaptivity in these models, allowing the decoder to dynamically adjust its internal logic based on the specific noise pattern of each received signal.
The Gated Hypernet Decoder
The core innovation of this method is the use of a hypernetwork—a "network that predicts another network." This allows the decoder to be input-dependent, adapting its strategy to the noise encountered in each packet.
As shown in Figure 1, the architecture consists of several key components:
- The Hypernetwork ($f$): A four-layer neural network that takes current messages in the decoding graph as input and predicts the weights ($\theta$) and gating values ($\sigma$) for the decoding nodes.
- The Message Generation Network ($h$): A two-layer network that performs the actual message passing, with its behavior dictated by the hypernetwork.
- Gating Mechanism: The hypernetwork produces a gating signal ($\sigma$) that scales activations, effectively allowing the model to re-route or "turn off" parts of the graph architecture on the fly.
- Damping Factor ($c$): To ensure training stability, a learnable damping factor acts as a residual connection, blending new neural-generated messages with those from previous iterations.
Experimental Results
The proposed decoder was evaluated on synthetic datasets using Polar codes in $(N=128, K=64)$ and $(N=32, K=16)$ configurations under Additive White Gaussian Noise (AWGN) channel models.
Key results from the performance analysis include:
- Matching the Gold Standard: For the $(32, 16)$ polar code at high Signal-to-Noise Ratios (SNRs), the gated hypernet matched the Bit-Error-Rate (BER) of the Successive List Cancellation (SLC) method. This is a significant milestone, as SLC is typically considered the performance upper bound for BP-based decoders.
- Outperforming Prior Neural Decoders: The method demonstrated a 0.1dB to 0.2dB improvement over previous state-of-the-art neural polar decoders, such as those proposed by Xu et al.
- Ablation Success: Ablation studies confirmed that the "Complete Method" (Hypernet + Gating + Damping) significantly outperforms versions missing any single component, highlighting the necessity of the damping factor for convergence.
Conclusion
By introducing input-dependent adaptivity through hypernetworks and gating, this work achieves state-of-the-art performance for neural polar decoders. While the computational overhead of running a hypernetwork per iteration is a consideration for hardware implementation, the ability to match SLC accuracy while maintaining the parallelizable nature of BP represents a substantial step forward in coding theory. Future work will focus on scaling these hypernetworks to larger block sizes and optimizing them for low-power hardware environments.
Summary updated