Chapter Summary

Chapter Summary

Key Points

• 1.

  NN OTFS receivers match or beat classical detectors. Pure NN (CNN or transformer): 0.2 dB worse than MP on idealized channels; 1-4 dB better on realistic channels with fractional Doppler, phase noise, HW imperfections. Gain grows with channel non-ideality.

• 2.

  Learned pilots give 3 dB MSE improvement. End-to-end trained pilot patterns (Ma-Wang-Caire 2020 CommIT) achieve 3 dB better channel estimation MSE on target profile. Multi-profile training: 2 dB advantage across diverse channels, 3% spectral efficiency gain over classical.

• 3.

  Model-based deep unfolding is the sweet spot. Unfolding an iterative algorithm ($\mathcal{A}$) into $T$ NN layers gives both classical structure (robustness, interpretability) and learning flexibility. 1-2 dB improvement over classical, with out-of-distribution robustness pure NN lacks.

• 4.

  Training: simulation-real gap is ~1 dB. Trained purely on simulation: 2-3 dB worse on real channels. Mitigate with domain randomization, fine-tuning, adversarial examples. Total gap: $\sim 0.5-1$ dB. Accept and deploy.

• 5.

  Federated learning preserves privacy. Train across $K$ UEs without centralized data. Convergence: $\mathcal{O}(K)$ extra rounds vs centralized. GDPR/CCPA-compliant. Essential for commercial 6G deployment.

• 6.

  Adversarial robustness is mandatory for safety. NN detectors susceptible to adversarial perturbations (V2X safety, etc.). Adversarial training: $2\times$ training cost, 0.5 dB clean penalty, $10\times$ better under attack.

• 7.

  Deployment pattern for 6G: vendor pre-training + federated personalization + online fine-tuning. Pre-trained NN detectors in UE chip, updated via OTA. Unfolded MP for safety-critical. Pure NN for data-heavy applications.

• 8.

  Standardization: 5G NR: vendor-proprietary AI/ML. 5G Advanced (Rel. 18): AI/ML framework. 6G Foundation (Rel. 20-21): native AI/ML, including learned pilots and unfolded detectors. 2030+: mainstream.