[Paper Review] TurboNet: A Model-driven DNN Decoder Based on Max-Log-MAP Algorithm for Turbo Code

Yunfeng He, Jing Zhang+2 more|arXiv (Cornell University)|May 25, 2019

Wireless Signal Modulation Classification12 references4 citations

TL;DR

TurboNet proposes a model-driven deep neural network (DNN) decoder for Turbo codes by unfolding the iterative max-log-MAP algorithm and parameterizing its components with learnable weights. Trained via a redefined loss function, TurboNet achieves superior bit error rate (BER) performance over traditional max-log-MAP decoding with fewer parameters and lower computational complexity, while generalizing well across diverse SNR conditions.

ABSTRACT

This paper presents TurboNet, a novel model-driven deep learning (DL) architecture for turbo decoding that combines DL with the traditional max-log-maximum a posteriori (MAP) algorithm. To design TurboNet, we unfold the original iterative structure for turbo decoding and replace each iteration by a deep neural network (DNN) decoding unit. In particular, the DNN decoding unit is obtained by parameterizing the max-log-MAP algorithm rather than replace the whole decoder with a black box fully connected DNN architecture. With the proposed architecture, the parameters can be efficiently learned from training data, and thus TurboNet learns to appropriately use systematic and parity information to offer higher error correction capabilities and decrease computational complexity compared with existing methods. Furthermore, simulation results prove TurboNet's superiority in signal-to-noise ratio generalizations.

Motivation & Objective

• To address the limitations of data-driven deep learning decoders, which suffer from high parameter counts and poor interpretability due to black-box architectures.
• To overcome the computational inefficiency and high dimensionality issues in training fully connected DNNs for long block-length Turbo codes.
• To develop a model-driven DNN architecture that integrates domain-specific knowledge of the max-log-MAP algorithm to improve decoding performance and generalization.
• To reduce the number of parameters and computational complexity compared to existing RNN-based or fully connected DNN decoders while maintaining or improving error correction capability.

Proposed method

• Unfold the iterative structure of the turbo decoder into a sequence of independent DNN decoding units, one per iteration.
• Parameterize the max-log-MAP algorithm by introducing learnable weights for the logarithmic operations in the Jacobian log-sum-exp approximation.
• Replace the traditional max-log-MAP computation with a differentiable DNN layer that learns optimal scaling and compensation factors for the LLR updates.
• Design a redefined loss function that directly minimizes the difference between predicted and target LLRs generated by the log-MAP algorithm with a fixed number of iterations.
• Train the network using stochastic gradient descent (SGD) and ADAM optimizer with a batch size of 500 and a learning rate of 1e-5.
• Use a single training SNR (0 dB) and a target iteration count (T=6) for log-MAP LLR generation to balance learning robustness and generalization.

Experimental results

Research questions

• RQ1 Can a model-driven DNN architecture that parameterizes the max-log-MAP algorithm outperform traditional max-log-MAP decoding in BER performance for Turbo codes?
• RQ2 Does the proposed TurboNet decoder generalize well across different SNR values despite being trained at a single SNR?
• RQ3 Can TurboNet achieve lower computational complexity and fewer parameters than data-driven RNN or fully connected DNN decoders while maintaining or improving performance?
• RQ4 How does the number of decoding units and the choice of training SNR and target iteration count affect the generalization and error correction capability of the DNN decoder?

Key findings

• TurboNet achieves lower bit error rate (BER) than the traditional max-log-MAP algorithm with five iterations across all SNR ranges for both (40,132) and (40,92) Turbo codes.
• For the (40,132) code, TurboNet outperforms the max-log-MAP algorithm with five iterations and matches the performance of the log-MAP algorithm with three iterations.
• For the punctured (40,92) code, TurboNet outperforms max-log-MAP with three iterations and achieves BER performance comparable to log-MAP with three iterations.
• TurboNet generalizes well across SNR values, maintaining superior performance even when tested at SNRs different from the training SNR (0 dB).
• The decoder has only 17.8K parameters, significantly fewer than the 3.85M parameters in the neural BCJR decoder from [3], while achieving faster inference time (1.39e-4s vs. 5.89e-3s).
• TurboNet reduces computational latency compared to max-log-MAP with five iterations, completing decoding in 1.39e-4s versus 2.3e-4s.