Learning During Detection: Continual Learning for Neural OFDM Receivers via DMRS

1. Problem Statement

Deep Neural Network (DNN) based receivers for Orthogonal Frequency-Division Multiplexing (OFDM) systems offer superior Bit Error Rate (BER) performance compared to traditional methods (e.g., Least Squares with Linear Minimum Mean Square Error) by learning complex channel correlations. However, they suffer from a critical limitation: distribution shift.

• The Issue: Communication channels change rapidly over time, frequency, and location. Models trained on static or simulated datasets often fail when deployed in real-world environments where channel statistics (e.g., delay spread, Doppler shift) deviate from the training distribution.
• Limitations of Existing Solutions:
  • Offline Retraining: Requires collecting new data and retraining, which is slow and impractical for dynamic channels.
  • Alternating Training/Inference: Conventional online fine-tuning methods often require dedicated pilot symbols or time intervals for training. This introduces high pilot overhead, interrupts user data transmission, and fails to track rapid channel variations occurring between updates.
  • Hybrid Receivers: Combining neural and traditional receivers improves robustness but is still constrained by the limitations of the traditional component under large distribution shifts.

2. Methodology

The paper proposes a zero-overhead online and continual learning framework where the neural receiver adapts during normal operation without interrupting inference or requiring extra resources.

A. Core Concept: Dual-Use Pilots

The central innovation is redesigning Demodulation Reference Signals (DMRS). Instead of treating pilots solely as inputs for demodulation or solely as labels for training, the proposed method uses them for both simultaneously.

• Mechanism: The transmitter randomizes pilot locations and symbol values. The receiver uses these pilots to:
  1. Demodulate: Treat them as known inputs to generate soft bits (LLRs) for data recovery.
  2. Train: Mask the corresponding pilot positions in the input tensor, treating the original pilot values as "ground truth" labels to compute loss and update weights.

B. Pilot Designs

Three flexible pilot configurations are introduced to balance backward compatibility and learning flexibility:

1. Fully Randomized: Pilots are scattered randomly across the resource grid with randomized constellation values. Maximizes diversity but complicates synchronization.
2. Hybrid: A subset of pilots is randomized for learning, while the rest follow conventional fixed patterns. Balances compatibility and adaptability.
3. Additional Pilots: Extra symbols are added alongside standard pilots specifically for learning. Compatible with fixed patterns but introduces minimal overhead.

C. Receiver Architectures

Two architectures are developed to exploit these pilot designs:

1. Architecture I (Parallel Design):
   • Uses two parallel paths: one for inference (unmasked pilots) and one for training (masked pilots).
   • A "clone" network is updated via backpropagation and its weights are periodically copied to the inference model.
   • Pros: Zero interruption to inference; high performance.
   • Cons: Higher computational complexity (requires maintaining two networks).
2. Architecture II (Forward-Pass Reusing):
   • Uses a single network. The forward pass generates LLRs for both data and masked pilots.
   • Backpropagation is applied directly to the same network after a mini-batch.
   • Pros: Lower computational complexity.
   • Cons: Requires brief pauses during backpropagation; fewer pilots available for demodulation.

D. Delay Analysis

The paper analyzes the latency introduced by backpropagation ($B_d$) relative to inference delay ($I_d$). It identifies three operating cases based on the ratio of transmission/decoding delays to inference time:

• Case I: Updates occur after every mini-batch (ideal).
• Case II: Updates occur every 2 mini-batches.
• Case III: Updates occur every 3 mini-batches (one forward, two backward passes).
  The analysis shows that even with less frequent updates (Cases II and III), the system can effectively track channel shifts without becoming a bottleneck.

3. Key Contributions

1. Zero-Overhead Strategy: A novel framework that enables continual learning without additional pilot overhead or service interruption by repurposing existing DMRS.
2. Flexible Pilot Designs: Introduction of three pilot structures (fully randomized, hybrid, additional) to support joint demodulation and learning under diverse constraints.
3. Dual-Architecture Approach: Development of two receiver designs (Parallel vs. Forward-Pass Reusing) offering a trade-off between computational resources and inference continuity.
4. Latency Characterization: A theoretical analysis of the delay introduced by model updates, defining the conditions under which online learning can keep pace with channel dynamics.
5. Empirical Validation: Extensive simulations demonstrating the framework's ability to recover from and track both slow and fast channel distribution shifts.

4. Simulation Results

The authors simulated an uplink OFDM system (64-QAM, 64 subcarriers) using the CoNet architecture (chosen for faster convergence over ResNet).

• Distribution Shift Vulnerability: A fixed neural receiver trained on a specific delay spread (40–50 ns) suffered catastrophic BER degradation when the delay spread shifted to [400, 410] ns, performing worse than traditional LMMSE receivers.
• Recovery Capability:
  • With 24k adaptation samples from the shifted distribution, the BER improved significantly.
  • With 50k samples, the receiver recovered to near-original performance levels.
• Continuous Tracking:
  • Slow Shifts: The model successfully tracked gradual changes in delay spread (e.g., 20 ns change every 1000 mini-batches), maintaining BER below $2 \times 10^{-3}$, while the fixed model degraded to $5 \times 10^{-2}$.
  • Fast Shifts: Even with rapid changes (20 ns every 480 samples), the online learning approach prevented severe degradation.
  • Random Shifts: Under random variations in delay spread, the adaptive model maintained low BER, whereas the fixed model failed.
• Update Frequency: Results showed that even with less frequent updates (Case III: update every 3rd mini-batch), the system converged effectively, though slightly slower than Case II.

5. Significance

This work addresses a fundamental barrier to the deployment of AI in wireless communications: the inability of static models to adapt to dynamic environments.

• Practicality: By eliminating the need for dedicated training pilots or service interruptions, the proposed method is highly feasible for real-world 5G/6G deployment.
• Robustness: It ensures that neural receivers do not become obsolete when channel conditions change, bridging the gap between the high performance of DL and the reliability requirements of wireless systems.
• Efficiency: It demonstrates that "learning during detection" is computationally viable and does not require massive increases in overhead, making it a scalable solution for future intelligent wireless networks.