DeepRx

Unlike traditional receivers that optimize each block independently, DeepRx trains all processing stages simultaneously using backpropagation. This holistic approach discovers synergistic signal representations that individual block optimization cannot achieve.

• Eliminates information bottlenecks between discrete processing blocks
• Learns to compensate for hardware impairments implicitly
• Adapts to non-linear channel effects that linear equalizers cannot handle
• Achieves superior performance on complex channel models like the 3GPP CDL and TDL profiles

DeepRx ingests raw complex baseband samples without requiring manual feature extraction or preprocessing. The network learns optimal representations directly from the in-phase and quadrature components.

• Operates on complex-valued tensors preserving phase information
• Eliminates the need for expert-designed feature engineering
• Handles variable-length input sequences for flexible frame structures
• Compatible with arbitrary sampling rates and bandwidth configurations

Traditional receivers require explicit pilot-based channel estimation and timing synchronization algorithms. DeepRx learns to perform these functions implicitly within its hidden layers, recovering transmitted symbols without dedicated reference signals.

• Reduces pilot overhead by learning from data symbols
• Robust to residual carrier frequency offset and sampling clock drift
• Maintains performance under high Doppler spread conditions
• Enables pilotless communication schemes for maximum spectral efficiency

DeepRx produces log-likelihood ratios (LLRs) for each transmitted bit, providing soft information to the channel decoder. The network learns non-linear decision boundaries that outperform classical maximum-likelihood demapping in the presence of hardware impairments.

• Models residual distortion from power amplifier non-linearity
• Accounts for I/Q imbalance and phase noise in the learned mapping
• Provides calibrated uncertainty estimates for iterative decoding
• Outperforms Gaussian-approximated LLRs on realistic channel models

DeepRx can incorporate known algorithmic structures as non-trainable layers, combining the efficiency of classical signal processing with the adaptability of neural networks. This model-based approach improves data efficiency and interpretability.

• Integrates Fast Fourier Transform (FFT) layers for OFDM processing
• Uses attention mechanisms to focus on relevant time-frequency resources
• Employs residual connections that mimic iterative interference cancellation
• Achieves strong performance with limited training data compared to pure black-box approaches

DeepRx architectures are designed for efficient inference on embedded hardware, using quantization-aware training and structured pruning to meet the latency and power constraints of real-time wireless systems.

• Supports INT8 quantization with minimal accuracy degradation
• Compatible with neural processing unit (NPU) acceleration
• Achieves sub-millisecond inference latency on 5G subcarrier grids
• Scales from massive MIMO base stations to IoT sensor nodes