DeepRx MIMO

Unlike conventional MIMO receivers that cascade separate spatial and temporal processing blocks, DeepRx MIMO applies joint spatio-temporal attention across all antenna elements and symbol periods simultaneously. This unified approach allows the model to learn optimal detection strategies that exploit correlations across both dimensions, mitigating inter-symbol interference and inter-stream interference in a single forward pass.

DeepRx MIMO replaces the entire traditional MIMO receiver chain—including channel estimation, equalization, and log-likelihood ratio (LLR) computation—with a single neural network trained end-to-end. The model ingests raw received IQ samples from all antennas and directly outputs soft bit estimates, learning to implicitly perform channel estimation and interference cancellation without explicit algorithmic decomposition.

The architecture employs complex-valued attention that operates natively on IQ baseband samples, preserving both magnitude and phase relationships. Key components include:

• Rotary Position Embedding (RoPE) adapted for complex-domain temporal encoding
• Multi-head self-attention across antenna streams
• Cross-attention between spatial and temporal feature representations This preserves the rich geometric structure of complex signals that would be lost with real-valued processing.

DeepRx MIMO scales efficiently to massive MIMO configurations with tens or hundreds of antenna elements. The transformer-based architecture processes antenna streams as a set of tokens, with computational complexity scaling gracefully through:

• Multi-head attention that parallelizes spatial processing
• Learned dimensionality reduction of the antenna dimension
• Shared attention weights across antenna groups This enables practical deployment in 5G and future 6G systems.

Rather than explicitly estimating the channel matrix and then performing detection, DeepRx MIMO learns to implicitly estimate and utilize channel state information within its hidden representations. The model is trained on diverse channel realizations and learns to extract sufficient statistics directly from pilot-contaminated received signals, often outperforming conventional methods that rely on explicit, imperfect channel estimates.

Training DeepRx MIMO with realistic hardware impairment models—including IQ imbalance, phase noise, and power amplifier non-linearity—enables the receiver to learn compensation strategies that are difficult to achieve with traditional model-based approaches. The unified architecture adapts its detection strategy to the specific impairment profile of the deployed hardware, improving practical link performance.

The foundational end-to-end neural receiver architecture that DeepRx MIMO extends. It replaces the entire traditional signal processing chain—channel estimation, equalization, and demapping—with a single deep learning model. DeepRx operates on single-antenna or single-stream signals, using a convolutional or transformer backbone to map received IQ samples directly to log-likelihood ratios (LLRs). DeepRx MIMO generalizes this concept to the multi-antenna case by adding spatial processing dimensions.

• Input: Raw IQ samples from a single receiver chain
• Output: Soft bit estimates (LLRs) for channel decoding
• Key innovation: Eliminates block-based algorithm partitioning in favor of a jointly optimized model

An attention mechanism that simultaneously models dependencies across both spatial dimensions (antenna elements) and temporal dimensions (symbol periods) in a multi-antenna signal. Unlike separate spatial and temporal processing stages, joint attention allows the model to learn correlations between what happens at different antennas at different times. This is a core enabler for DeepRx MIMO, as it captures the full structure of the space-time channel matrix in a unified operation.

• Spatial axis: Antenna ports, beam indices, or spatial streams
• Temporal axis: Symbol periods or time-domain samples
• Advantage: Learns cross-antenna interference patterns that vary over time

A transformer-based model designed to process channel state information (CSI) matrices directly. It leverages self-attention to capture spatial and frequency correlations within the channel response. In a DeepRx MIMO context, a CSI Transformer can serve as a pre-processing stage or an integrated component that provides learned channel representations to the detection network, replacing traditional minimum mean square error (MMSE) or least squares estimators.

• Input: CSI matrices across antennas and subcarriers
• Output: Refined channel estimates or compressed CSI feedback
• Relevance: Provides high-quality spatial signatures for MIMO detection

A transformer network that predicts optimal beamforming weights directly from channel state information or received signal snapshots. It replaces traditional optimization algorithms like singular value decomposition (SVD) or iterative water-filling with a learned attention-based mapping. In a DeepRx MIMO system, a Beamforming Transformer can be jointly trained with the neural receiver to optimize the entire beamforming-plus-detection pipeline end-to-end.

• Input: CSI or raw antenna snapshots
• Output: Complex beamforming weight vectors
• Integration: Can be co-designed with DeepRx MIMO for joint optimization

An extension of the standard attention mechanism that operates natively on complex numbers, preserving the magnitude and phase relationships inherent in IQ baseband signals. Standard attention operates on real-valued vectors, requiring complex signals to be split into real and imaginary components, which can destroy phase coherence. Complex-valued attention uses complex-valued weights and activations, making it particularly well-suited for DeepRx MIMO where phase alignment across antennas is critical for spatial multiplexing.

• Preserves: Phase relationships between antenna streams
• Operations: Complex-valued matrix multiplications and softmax
• Benefit: More expressive processing of coherent multi-antenna signals

A neural network based on the transformer architecture that performs channel equalization by attending to a sequence of received symbols to mitigate inter-symbol interference (ISI) and recover transmitted data. In a MIMO context, the Transformer Equalizer can be extended to handle spatial multiplexing interference by attending across both time and antenna dimensions. DeepRx MIMO often incorporates equalization implicitly within its unified architecture rather than as a separate block.

• Function: Reverses channel distortion and inter-stream interference
• Mechanism: Self-attention over received symbol sequences
• MIMO extension: Cross-antenna attention for spatial interference cancellation