[Paper Review] All you need is feedback: Communication with block attention feedback codes

Emre Özfatura, Yulin Shao+3 more|arXiv (Cornell University)|Jun 19, 2022

Wireless Signal Modulation Classification4 citations

TL;DR

This paper proposes Generalized Block Attention Feedback (GBAF) codes, a deep learning-aided feedback channel coding scheme using a transformer-based architecture with block-by-block processing and attention mechanisms. GBAF achieves order-of-magnitude lower block error rates (BLER) than prior DNN-based codes, especially in low SNR and noisy feedback regimes, with design C (ReLU-activated) offering optimal trade-off between performance and computational efficiency.

ABSTRACT

Deep learning based channel code designs have recently gained interest as an alternative to conventional coding algorithms, particularly for channels for which existing codes do not provide effective solutions. Communication over a feedback channel is one such problem, for which promising results have recently been obtained by employing various deep learning architectures. In this paper, we introduce a novel learning-aided code design for feedback channels, called generalized block attention feedback (GBAF) codes, which i) employs a modular architecture that can be implemented using different neural network architectures; ii) provides order-of-magnitude improvements in the probability of error compared to existing designs; and iii) can transmit at desired code rates.

Motivation & Objective

• Address the challenge of ultra-reliable short-packet communication in noisy, finite-blocklength regimes where conventional codes fall short.
• Overcome limitations of human-crafted feedback codes—such as sensitivity to quantization and finite precision—by leveraging end-to-end deep learning.
• Design a flexible, modular feedback coding architecture that supports variable coding rates and reduces communication overhead via block-by-block processing.
• Achieve superior block error rate (BLER) performance compared to existing DNN-based feedback codes, especially in low signal-to-noise ratio (SNR) and noisy feedback conditions.
• Optimize the neural architecture for practical deployment by balancing performance, computational cost, and robustness to feedback channel impairments.

Proposed method

• Propose GBAF codes as a generalized, modular feedback coding framework using sequence-to-sequence encoding with block-by-block message processing.
• Employ a transformer-based architecture with multi-head self-attention and feedforward networks to model long-range dependencies in feedback interactions.
• Integrate a belief network to maintain and update the receiver’s posterior estimate of the transmitted message across feedback rounds.
• Use end-to-end unsupervised training to minimize reconstruction error at the receiver, treating the wireless channel as a stochastic layer.
• Design multiple feature extractor variants (C and D) to evaluate trade-offs between performance and computational efficiency.
• Compare activation functions ReLU and GeLU in the feature extractor to assess their impact on BLER performance under noiseless and noisy feedback conditions.

Experimental results

Research questions

• RQ1 Can a deep learning-based feedback code achieve significantly lower block error rates than existing DNN-based and classical feedback codes in low SNR regimes?
• RQ2 How does the choice of neural network architecture—specifically, different feature extractor designs—affect the performance and computational cost of feedback codes?
• RQ3 What is the impact of feedback channel quality (noiseless vs. noisy) on the performance of GBAF codes, and which architecture is most robust under degradation?
• RQ4 How do different activation functions (ReLU vs. GeLU) influence the BLER performance of GBAF codes, especially at ultra-low error rates?
• RQ5 Can the proposed GBAF framework support flexible coding rates and reduce communication overhead through block-by-block processing compared to full-sequence feedback schemes?

Key findings

• GBAF codes achieve a block error rate (BLER) as low as 5×10⁻⁹ at -1 dB feedforward SNR, significantly outperforming AttentionCode and other prior DNN-based codes.
• With noiseless feedback, GBAF using design C achieves a 1.2 dB SNR gain over AttentionCode to reach a BLER of 10⁻⁸.
• Design C (ReLU-activated) outperforms design D in noisy feedback conditions (20 dB feedback SNR), maintaining low BLER while being twice as computationally efficient.
• GeLU activation improves BLER performance by 5× at -1 dB SNR compared to ReLU under noiseless feedback, but ReLU becomes superior when feedback SNR degrades to 20 dB.
• When both feature extractor and belief network are used, GeLU yields lower BLER than ReLU in noiseless feedback, but ReLU is more robust under noisy feedback, confirming its practical advantage in real-world conditions.
• The default GBAF configuration is set to design C with ReLU activation, offering the best balance of performance, robustness to feedback noise, and computational efficiency.