DeepRx

The entire receiver is a single differentiable function mapping received IQ samples directly to log-likelihood ratios (LLRs) or bit estimates. This allows joint optimization of all processing stages simultaneously using stochastic gradient descent, eliminating the information bottlenecks inherent in modular, block-based designs. The loss function directly targets bit error rate or mutual information, ensuring the learned algorithm optimizes for the true end goal rather than intermediate proxy metrics like minimum mean squared error.

DeepRx does not rely on explicit algorithms like Least Squares (LS) or Linear Minimum Mean Square Error (LMMSE) estimators. Instead, a convolutional or transformer backbone implicitly learns to:

• Estimate the channel from pilot symbols embedded in the received grid.
• Interpolate the channel response across time and frequency.
• Equalize the signal by inverting channel distortions. This learned approach is particularly robust in high-Doppler or low-pilot-density scenarios where traditional model-based methods fail.

The original DeepRx architecture employs a deep residual network (ResNet) of 1D convolutional layers operating along the frequency dimension of an OFDM resource grid. Key design choices include:

• Dilated convolutions to expand the receptive field and capture inter-subcarrier interference.
• Skip connections to stabilize training and enable gradient flow through very deep networks.
• Input processing that concatenates raw received symbols with known pilot symbols, treating the channel estimation task as an inpainting problem.

Modern iterations replace or augment the convolutional backbone with self-attention mechanisms to better model long-range dependencies across the time-frequency resource grid. A DeepRx Transformer tokenizes the received OFDM grid into a sequence of subcarrier or time-frequency patch tokens, then applies multi-head self-attention. This allows the model to dynamically weigh the importance of distant pilots and symbols, yielding superior performance in high-mobility channels with severe Doppler spread.

DeepRx produces soft-decision outputs in the form of log-likelihood ratios (LLRs) for each coded bit, enabling seamless integration with a downstream soft-input channel decoder like LDPC or Polar codes. The final layer of the network is a softmax or a learned LLR mapping that is trained to maximize the mutual information between the transmitted bits and the output LLRs, directly optimizing the coded block error rate (BLER) rather than uncoded symbol error rate.

The architecture extends naturally to multi-antenna systems. DeepRx MIMO processes the received signal across all antenna elements jointly, learning to perform spatial demultiplexing and interference suppression without explicit channel matrix inversion. The model can learn non-linear detection strategies that outperform traditional MMSE-IRC receivers, especially in overloaded MIMO scenarios where the number of spatial layers approaches or exceeds the number of receive antennas.