Computationally Efficient Neural Receivers via Axial Self-Attention

TL;DR

Under non-line-of-sight conditions, the proposed axial neural receiver outperforms global self-attention and convolutional neural receiver baselines at 10% BLER and 1% BLER respectively, with reduced computational complexity.

Abstract

Deep learning-based neural receivers offer promising physical-layer solutions for next-generation wireless systems. We propose an axial self-attention transformer neural receiver that achieves state-of-the-art Block Error Rate (BLER) performance with significantly improved computational efficiency during inference and large-scale training. By factorizing attention operations along temporal and spectral axes, the proposed architecture reduces computational complexity from $O((TF)^2)$ to $O(T^2F+TF^2)$, yielding substantially fewer floating-point operations and attention matrix multiplications per transformer block. Experimental validation under 3GPP Clustered Delay Line (CDL) channels demonstrates consistent performance gains across varying mobility scenarios. Under non-line-of-sight conditions, our proposed axial neural receiver outperforms global self-attention and convolutional neural receiver baselines at 10% BLER and 1% BLER respectively, with reduced computational complexity.

Figures (4)

• Figure 1: Architecture of axial attention transformer-based neural receiver. It comprises a 2D convolutional input projection, 2D learned positional encoding, six transformer blocks, and a 2D convolutional output projection.
• Figure 2: Axial transformer block with sequential time-axis and frequency-axis multi-head attention preceded by layer normalization. Factorized attention operations reduce computational complexity while maintaining long-range dependency modeling through residual connections.
• Figure 3: BLER performance under -C and -D channels at user velocities 10--40 m/s.
• Figure 4: Model complexity of neural receiver architectures