Lightweight PCGAE-Net: Parallel CrossGate Attention and Bottleneck AutoEncoder for Efficient 5G Channel Prediction

Researchers introduce Lightweight PCGAE-Net, a new neural network architecture that reduces 5G channel prediction model size by 58% while improving accuracy by up to 6.0dB. The model addresses architectural inefficiencies in existing transformers through parallel attention mechanisms and a bottleneck autoencoder, enabling deployment on base-station hardware with computational constraints.

The paper tackles a critical infrastructure challenge in 5G networks: deploying accurate channel state information (CSI) predictors on resource-constrained base-station hardware. Current state-of-the-art models exceed 30 million parameters, making them impractical for edge deployment despite their superior predictive capabilities. Rather than applying post-hoc compression techniques like quantization or pruning, the researchers identify and fix two fundamental architectural flaws in existing models.

The first innovation involves restructuring attention mechanisms to eliminate dependency ordering bias. Traditional approaches apply temporal attention sequentially after spatial attention, which constrains what temporal patterns can be extracted. By routing both attention types through parallel paths to the same normalized input and combining outputs via a learned CrossGate module, the model captures richer feature interactions independently. The second key contribution replaces expensive full self-attention at the bottleneck with a lightweight autoencoder using 1x1 convolutions, reducing computational cost quadratically while maintaining information integrity through auxiliary reconstruction loss.

These architectural improvements enable dramatic model compression—from 30M+ to 8.54M parameters—while simultaneously achieving performance gains of 3.26dB to 6.0dB depending on channel velocity conditions. This represents a rare scenario where model efficiency and accuracy improve simultaneously rather than trading off against each other. The practical implications are substantial for 5G operators seeking to implement intelligent beamforming and resource management on existing hardware infrastructure without significant capital investment in equipment upgrades.

• →Model size reduced by 58% while accuracy improved by up to 6.0dB through architectural optimization rather than post-hoc compression
• →Parallel attention mechanisms eliminate sequential dependency bias that constrains temporal pattern learning
• →Bottleneck autoencoder with auxiliary loss prevents information collapse while reducing computational complexity quadratically
• →8.54M parameter model enables deployment on resource-constrained base-station hardware for real-time 5G channel prediction
• →Architecture improvements demonstrate efficiency gains need not sacrifice accuracy in deep learning model design