Neural Networks based Multiple Access (NeNeMa): An AI Deep Learning Approach for Beyond-Orthogonality Spectral Efficiency

This paper proposes Neural Networks based Multiple Access (NeNeMa), a groundbreaking multiple access scheme that leverages deep learning to enable intentional, separable signal superposition in wireless communications. Unlike traditional orthogonal multiple access (OMA) and conventional powerdomain NOMA (PD-NOMA), NeNeMa employs neural encoders and decoders to learn optimal signal representations that allow multiple users to simultaneously utilize the entire time-frequency spectrum while maintaining separability at intended receivers. The system architecture integrates transformer-based attention mechanisms, variational autoencoders for latent space optimization, and adversarial training for robustness. Theoretical analysis establishes the fundamental limits of NeNeMa through novel information-theoretic bounds, demonstrating that it achieves the capacity region of the Gaussian multiple access channel under learned signal representations. Simulation results show that NeNeMa achieves spectral efficiency gains of 3.8-5.2× compared to OFDMA and 2.1-3.7× compared to PD-NOMA in realistic multipath fading channels, while providing inherent physicallayer security with 15-25 dB advantage in secrecy capacity over conventional schemes. The proposed scheme eliminates the need for successive interference cancellation (SIC) and offers a practical path toward realizing beyond-orthogonality communications for 6G networks. Implementation considerations, including computational complexity, training overhead, and hardware feasibility, are thoroughly analyzed, demonstrating practical deployment viability for next-generation wireless systems.