DeepRx

The entire receiver pipeline—from channel estimation and equalization to demapping and decoding—is a single differentiable function. This allows gradients to flow from the bit-error loss back through every stage, enabling joint optimization of all components simultaneously. Unlike modular designs where each block is optimized in isolation, DeepRx discovers co-adapted internal representations that maximize end-to-end performance.

DeepRx does not contain an explicit channel estimation module. Instead, the network learns to implicitly extract channel state information from the received pilot and data symbols as an internal latent representation. This eliminates the need for explicit interpolation of pilot responses and avoids the performance ceiling imposed by classical estimators like Least Squares (LS) or Minimum Mean Square Error (MMSE) in high-mobility or low-pilot-density scenarios.

Rather than computing Log-Likelihood Ratios (LLRs) using the classical assumption of Gaussian noise and fixed constellation grids, DeepRx learns a direct mapping from received soft symbols to bit probabilities. This is particularly advantageous under non-linear hardware impairments or non-Gaussian interference, where the true decision boundaries deviate significantly from the idealized Voronoi regions assumed by traditional soft demappers.

Because DeepRx is trained on data that includes real-world power amplifier non-linearities, oscillator phase noise, and I/Q imbalance, the network learns to compensate for these distortions intrinsically. This contrasts with classical receivers that require separate, often complex, calibration and compensation algorithms for each impairment. The result is a receiver that maintains high throughput even on lower-cost, less-linear RF front-ends.

DeepRx can be trained to operate effectively with non-standard or sparse pilot patterns that would severely degrade classical channel estimators. The network learns the optimal interpolation strategy for any given reference signal configuration directly from data. This enables network operators to reduce pilot overhead and reclaim time-frequency resources for data transmission, directly increasing spectral efficiency.

DeepRx architectures typically employ complex-valued neural networks where weights, biases, and activations are complex numbers, and backpropagation uses Wirtinger calculus. This preserves the phase information inherent in I/Q samples, which is critical for coherent detection. Real-valued architectures that split I and Q into separate channels lose the algebraic structure of complex multiplication, degrading the network's ability to model phase rotations and frequency offsets.

A hybrid methodology that bridges classical algorithms and deep learning. An iterative optimization algorithm, such as ISTA or ADMM, is unrolled into a neural network where each layer corresponds to one iteration. Learnable parameters replace hand-crafted step sizes and thresholds. This contrasts with DeepRx's fully learned 'black-box' approach by retaining the structural priors of the original algorithm, leading to improved interpretability, faster training, and stronger generalization guarantees with fewer parameters.

A deep learning model that replaces traditional linear or decision-feedback equalizers to invert the dispersive effects of a wireless channel. Key capabilities include:

• Handling severe non-linear distortions and inter-symbol interference
• Learning complex channel inversions that classical filters cannot model
• Operating as a standalone component, unlike DeepRx which absorbs equalization into the unified receiver
• Often implemented with CNN or RNN architectures for time-series signal recovery

A hybrid model-based deep learning architecture that integrates the classical Kalman filter's structural flow with small neural networks. These networks learn the unknown system dynamics and noise statistics directly from data. For channel tracking in high-mobility scenarios, KalmanNet provides robust state estimation without requiring an explicit mathematical model of the channel evolution. This contrasts with DeepRx, which learns the entire reception task without incorporating explicit Bayesian filtering priors.

A neural network architecture where weights, biases, and activations are complex numbers, with backpropagation performed using Wirtinger calculus. This inherently preserves the phase information critical for coherent wireless signal processing. DeepRx implementations often leverage complex-valued layers to directly process I/Q baseband samples without separating real and imaginary components, enabling the network to learn phase-sensitive transformations essential for carrier synchronization and coherent detection.

The use of deep neural networks to learn the mapping from received pilot signals to the wireless channel response. This replaces or augments classical estimators like Least Squares (LS) or Minimum Mean Square Error (MMSE). In a traditional receiver, channel estimation is a distinct preprocessing step. DeepRx subsumes this function entirely—the network learns an implicit representation of the channel state directly from raw received samples, eliminating the need for an explicit estimation module.