Introduction
Machine learning (ML) models for wireless communication are typically trained using standard classification losses, such as Binary Cross-Entropy (BCE). However, this approach is often suboptimal for real-world coded systems. BCE treats bits as independent entities, ignoring the block-based structure of Forward Error Correction (FEC) codes and over-prioritizing noisy, low-SNR signals. This paper introduces "block-aware" loss functions and an "SNR deweighting" technique to bridge the gap between bit-level optimization and block-level reliability.
The BER-BLER Gap and SNR Imbalance
In modern communications, data is transmitted in blocks protected by FEC. A single bit error usually renders the entire block invalid, making Block Error Rate (BLER) the primary performance metric. Standard BCE loss minimizes Bit Error Rate (BER) by learning posterior marginals, which does not necessarily minimize the probability of block failure.
Furthermore, training models across a wide range of Signal-to-Noise Ratios (SNRs) often leads to an imbalance. Low-SNR samples produce significantly larger loss values, causing the optimizer to focus on these noisy signals while neglecting high-SNR regimes where the system should be most reliable. This results in a performance "floor" or even a breakdown at high SNRs.
Proposed Methods
Block-Aware Loss Functions
To align training objectives with BLER, the authors propose losses that penalize bit errors jointly:
- Product Loss: Represents the probability that all bits in a block are correctly decoded.
- SmoothMax & LogSumExp: These functions act as differentiable approximations of the maximum operator, focusing training on the most uncertain bit in a block—the one most likely to cause a block failure.
SNR Deweighting
To address SNR imbalance, an adaptive weighting scheme is introduced. Training batches are partitioned by SNR, and the cumulative loss for each level is tracked. Samples are then weighted by the inverse of their cumulative loss in subsequent epochs. This ensures that the model maintains precision in high-SNR conditions without being overwhelmed by low-SNR noise.
Experimental Evaluation
The methods were evaluated using high-fidelity simulations in the NVIDIA Sionna library across two scenarios:
- 5G MIMO-OFDM: A complex system utilizing a Deep Unfolded Interleaved Detection and Decoding (DUIDD) receiver with 5G LDPC codes.
- SISO AWGN: A single-input single-output channel used to verify the generalizability of the proposed losses.
Key Results
The proposed interventions yielded significant performance gains without increasing model complexity.
As shown in Figure 4, using BLER-specific loss functions in the 5G MIMO-OFDM scenario improved performance by 0.62 dB at a 1% BLER target compared to standard BCE. While BCE and MSE are effective for BER, they fall short in BLER performance.
SNR deweighting provided a 0.54 dB improvement over naive training. Crucially, as illustrated in Figure 6, it prevents the performance degradation at high SNRs that occurs when the model over-optimizes for low-SNR conditions.
Conclusion and Future Work
This research demonstrates that the choice of loss function is a critical design parameter in ML-assisted communications. By adopting block-aware objectives and SNR deweighting, practitioners can achieve "free" reliability gains. Future work will address potential gradient challenges in very large blocks and explore the hardware implications of these optimized receivers.
Summary updated