Abstract
This Special Issue of the IEEE Journal on Emerging and Selected Topics in Circuits and Systems (JETCAS) explores the critical synergy between Artificial Intelligence (AI) and the evolving landscape of 5G and Beyond 5G (B5G) networks. By focusing on implementations, algorithms, and optimizations, the featured research moves beyond theoretical modeling to address the practical, hardware-constrained realities of modern telecommunications engineering.
Introduction: Addressing Wireless Uncertainties
As the industry transitions from 4G to 5G and B5G, wireless networks face "extreme requirements," including massive throughput, ultra-low latency, and the capacity to connect an unprecedented density of devices. Traditional mathematical models often struggle to account for the "wireless uncertainties" of these complex environments—especially when a priori knowledge of the channel is incomplete or constantly changing.
While AI has demonstrated transformative potential in fields like computer vision, its application to the physical and circuit layers of telecommunications is still maturing. This special issue addresses the gap between high-level AI theory and the rigorous demands of hardware implementation, showcasing eight technical papers that provide concrete solutions for baseband processing, hardware-efficient learning, and intelligent wireless sensing.
Methodological Themes in AI-Driven Communications
The research presented is categorized into three primary methodological pillars:
1. Learning for Baseband Processing
Advanced signal processing is critical for managing the complexities of B5G. Key innovations include:
- Neural Codebooks: Researchers have developed deep learning models that function as general "neural codebooks" for non-orthogonal precoding in millimeter-wave (mmWave) communications. This approach significantly reduces the overhead of retraining when channel conditions shift.
- Unsupervised Learning via Syndrome Analysis: By implementing "syndrome-enabled" unsupervised learning for polar decoders, systems can learn and adapt without the need for massive labeled datasets, which are often impractical to obtain in real-time scenarios.
- Deep Learning-Aided Belief Propagation (BP): Traditional BP decoders are optimized using deep learning to achieve high-performance error correction with significantly lower computational complexity.
2. Hardware Implementations and Prototyping
AI algorithms must be translated into efficient silicon to be viable for infrastructure deployment.
- Specialized ASIC Design: The issue highlights Application-Specific Integrated Circuit (ASIC) designs for neural self-interference (SI) cancellation and high-speed polar decoders.
- SimuNN Framework: To accelerate development, the SimuNN framework provides a pre-RTL (Register-Transfer Level) simulation environment for early-phase verification and prototyping of neural networks before they are committed to silicon.
3. Wireless Sensing and Environmental Intelligence
Beyond data transmission, B5G networks serve as intelligent sensors for their environment.
- Transfer Learning for Massive MIMO: Semi-supervised automatic modulation classification (AMC) using convolutional auto-encoders (CAE) is applied to handle the scale of massive MIMO systems.
- Memristive Reservoir Computing: Deep Echo State Networks (DESN) utilizing memristive synapses offer a high-speed, low-error architecture for symbol detection in 5G networks.
Key Performance Gains and Insights
The synergy of AI and hardware-software co-design yields significant quantitative improvements over traditional methods:
- Energy Efficiency: AI-driven 5G positioning demonstrates a 47x to 85x improvement in energy efficiency compared to traditional assisted-GPS for continuous tracking.
- Interference Management: A neural hardware canceller achieved a 7 dB improvement in self-interference cancellation while utilizing only ~12% of the area and ~13% of the power required by traditional polynomial-based methods.
- Error Reduction: The DESN architecture showed a 10.31x reduction in prediction error compared to state-of-the-art neural designs, while unsupervised blind equalizers outperformed standard non-blind MMSE equalizers by 1.3 dB.
Conclusion and Future Outlook
The guest editors emphasize that while these results are promising, the field is in its early stages. Future research must focus on refining AI models to handle extreme wireless uncertainties, identifying technical hurdles for integrating AI-driven circuits into standardized infrastructure, and further optimizing device localization to reduce reliance on power-hungry satellite systems.
Editorial Leadership
Guest Editor Source, p. 3 Guest Editor Source, p. 3
Summary updated