DeepRx: Deep Learning Receiver for 5G

• DeepRx is a fully convolutional deep learning receiver for OFDM-based wireless links, combining joint channel estimation, equalization, and soft-bit demapping in a 5G-compliant framework.
• It leverages a ResNet-inspired architecture with dilated, depthwise separable convolutions to operate over the full time-frequency grid, optimizing performance in pilot-sparse and high Doppler scenarios.
• Extensions of DeepRx include MIMO detection, superimposed DMRS reception, influence-guided fine-tuning, and energy-aware model compression strategies to bridge research and practical deployments.

DeepRx is a fully convolutional deep learning receiver for OFDM-based wireless links, designed to be 5G NR-compliant and trained to map frequency-domain received symbols to bitwise soft outputs suitable for LDPC decoding. In its original formulation, DeepRx jointly replaces most of the conventional digital receiver chain—especially channel estimation, equalization, and soft demapping—with a single neural network operating over the full time-frequency resource grid; in later work, the same receiver family is analyzed as a fully convolutional DNN-based wireless receiver and as a multi-label classifier that predicts thousands of binary outcomes in parallel from received frequency-domain signals (Honkala et al., 2020, Tuononen et al., 19 Sep 2025).

1. System formulation and receiver role

DeepRx was introduced for a 5G uplink PUSCH link generated with MATLAB 5G Toolbox, in a SIMO uplink setting with 1 transmit antenna, 2 receive antennas, 14 OFDM symbols per TTI, 15 kHz subcarrier spacing, 26 PRBs $=312$ subcarriers, and 1 ms TTI duration. The received FFT-domain signal is modeled as

$$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$

where $i$ indexes OFDM symbols, $j$ indexes subcarriers, $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$, $x_{ij}\in\mathbb{C}$, $\mathbf{H}_{ij}\in\mathbb{C}^{N_r\times 1}$, and $\mathbf{n}_{ij}\in\mathbb{C}^{N_r\times 1}$ (Honkala et al., 2020).

The conventional receiver baseline consists of pilot-based channel estimation, interpolation of channel estimates over the resource grid, LMMSE equalization, soft demapping to LLRs, and LDPC decoding. DeepRx replaces steps 1–4 with one learned network; timing synchronization and OFDM demodulation, including CP removal and FFT, remain conventional. In consequence, DeepRx is not a raw-waveform receiver. It is a frequency-domain learned receiver front-end whose output is transparent to HARQ because it only changes the receiver front-end producing LLRs (Honkala et al., 2020).

The input construction is central. Three complex-valued tensors are formed: the received signal

$\mathbf{Y}\in\mathbb{C}^{S\times F\times N_r},$

the pilot tensor

$\mathbf{X}_p\in\mathbb{C}^{S\times F},$

and a raw pilot-only estimate

$$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$0

These are stacked as

$$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$1

and then split into real and imaginary parts to produce a real-valued input

$$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$2

For $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$3, the final real input has 10 channels. This gives the network aligned access to noisy received symbols, pilot mask and values, and raw pilot-based channel evidence over the whole TTI (Honkala et al., 2020).

The output is bitwise soft information. DeepRx directly emits bit-level logits $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$4, interpreted as LLR-like soft bits, rather than first producing equalized constellation symbols. This receiver formulation is one reason the model can remain 5G-compliant while replacing multiple conventional blocks at once (Honkala et al., 2020).

2. Architecture, output semantics, and training objective

DeepRx is a 2D fully convolutional network over the time-frequency resource grid. It preserves full $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$5 resolution throughout, outputs one prediction per resource element and per bit, avoids pooling and striding because every symbol needs a soft output, and has inference complexity that scales as

$$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$6

with respect to TTI dimensions, assuming pilot density stays roughly constant (Honkala et al., 2020).

The architecture uses a preactivation ResNet with batch normalization, residual connections, depthwise separable convolutions, and dilated convolutions. In the main architecture, the input convolution is $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$7 with 64 channels, followed by 11 ResNet blocks with depthwise separable convolutions, dilations increasing up to $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$8, widths increasing to 256 channels and then decreasing, and a final $$\mathbf{y}_{ij} = \mathbf{H}_{ij} x_{ij} + \mathbf{n}_{ij},$$9 convolution to $i$0 outputs per resource element. The resulting output tensor is

$i$1

and the main model has about 1.2M parameters (Honkala et al., 2020).

The network is trained as binary classification per bit using sigmoid cross-entropy, while the raw logits are used as soft values for decoding. The bit probability estimate is

$i$2

and the loss is

$i$3

The paper does not provide a more formal probabilistic calibration proof that the logits are exact LLRs; instead, it treats them as LLR-like soft outputs whose quality is validated through LDPC decoding performance (Honkala et al., 2020).

A further architectural feature is modulation-agnostic hierarchical bit mapping. The same output bit positions correspond to the same coarse-to-fine partition of the constellation regardless of modulation order, so one network can support multiple QAM orders by using output size corresponding to the highest supported modulation and masking only the active bits during loss computation or inference. This is presented as a practical 5G-compatibility mechanism (Honkala et al., 2020).

3. Empirical behavior and the interpretation of its gains

The original DeepRx results are framed against two LMMSE baselines: a practical receiver with LS channel estimation, interpolation, LMMSE equalization, and approximate demapping, and an LMMSE receiver with full channel knowledge. In the no-interference setting, DeepRx clearly outperforms the practical LMMSE receiver. With one pilot symbol, DeepRx nearly matches the LMMSE receiver with full channel knowledge; at high SINR, DeepRx with one pilot beats the practical LMMSE with two pilots by about a factor of 10 in BER; and in coded BER, the gain over practical LMMSE with two pilots is about 2 dB (Honkala et al., 2020).

The reported robustness is strongest in precisely the regimes where conventional blockwise processing is stressed. Under inter-cell interference, DeepRx outperforms even the known-channel LMMSE baseline, and in coded BER under interference it exceeds the known-channel benchmark by about 2 dB. Under high Doppler, practical LMMSE with one pilot degrades rapidly, practical LMMSE with two pilots remains reasonable up to around 400 Hz, and DeepRx with one pilot matches the known-channel LMMSE even at 500 Hz, corresponding to about 135 km/h UE velocity at 4 GHz (Honkala et al., 2020).

The paper’s interpretation of these gains is explicit. DeepRx appears to learn to exploit the known pilot symbols, the unknown data symbols’ constellation structure, and the local symbol distribution in the time-frequency grid jointly. Evidence comes from a restricted CNN receiver that is forced to estimate the channel from pilots only and then process data symbol-by-symbol using $i$4 convolutions. That restricted model yields only small gains over LMMSE, suggesting that the main benefit is not merely better pilot interpolation but joint use of pilots and data over the whole TTI (Honkala et al., 2020).

This interpretation is sharpened by manipulated symbol-distribution experiments. For QPSK with unusual but still phase-structured local distributions, DeepRx remains strong. For 16-QAM with an artificially altered spatial symbol distribution, DeepRx performs poorly and can be beaten by the restricted CNN. The paper reports average BER under this manipulation of $i$5 for the most significant two bits and $i$6 for the least significant two bits, and interprets the result as evidence that DeepRx has learned to rely on local data distribution for amplitude normalization and tracking. A plausible implication is that DeepRx’s strongest inductive bias is not generic “deep learning for channel estimation,” but implicit blind-aided channel inference coupled to local constellation statistics (Honkala et al., 2020).

The same paper is also explicit about limits. It does not provide hardware-level complexity, latency, or power comparisons; performance may depend on matching training and deployment distributions; and the internal learned mechanism is not fully interpretable. These caveats became central in later DeepRx work (Honkala et al., 2020).

4. Structural extensions: MIMO DeepRx and superimposed-DMRS DeepRx

DeepRx was extended from SIMO to SU-MIMO in “DeepRx MIMO: Convolutional MIMO Detection with Learned Multiplicative Transformations” (Korpi et al., 2020). The paper argues that naively scaling the original CNN by concatenating received signals and channel estimates is not enough, because MIMO spatial separation requires explicit support for multiplicative interactions between signal and channel. The proposed architecture therefore inserts a trainable transformation layer before the main DeepRx CNN, with two alternatives: an MRC-based transformation and a fully learned multiplicative transformation (Korpi et al., 2020).

In the MIMO version, the main DeepRx remains a CNN with 11 ResNet blocks and depthwise-separable 2D convolutional layers, following the same architecture as the earlier work but with quadruple channel count, capped at 512. The MIMO task is 4-layer SU-MIMO with 16 receive antennas in a 5G uplink shared channel simulation. Both transformation-layer variants clearly outperform the practical LMMSE benchmark, especially when pilots are sparse, and even with only one pilot symbol per TTI, DeepRx MIMO can nearly match LMMSE with perfect channel knowledge (Korpi et al., 2020).

A different extension appears in the SI-DMRS study for spectrally efficient 6G uplink multi-user OFDM. There, DeepRx is adapted to the much harder case where demodulation reference signals are superimposed on data over the full time-frequency grid rather than placed on dedicated orthogonal resource elements. The resulting receiver is a hybridized DeepRx: it begins with an SI-DMRS-aware LS channel estimate, refines that estimate with a Channel Estimation CNN, uses two parallel equalizers, applies a Detector CNN to refine equalized symbols, and maps the resulting features to LLRs with a Demapper CNN. The main architecture uses 12 ResNet blocks in the Channel Estimation CNN, 12 ResNet blocks in the Detector CNN, and 4 ResNet blocks in the Demapper CNN (Rezaie et al., 25 Jun 2025).

This SI-DMRS extension is important because it changes what DeepRx is solving. Under superimposed pilots, channel estimation is contaminated by self data interference, cross-user data interference, and residual pilot interaction, so the problem is no longer cleanly separable into pilot-only estimation followed by equalization. In the reported link-level evaluations, DeepRx outperforms conventional receivers for both SI-DMRS and orthogonal DMRS, and is especially advantageous for SI-DMRS at moderate-to-high SNR, where the reclaimed pilot overhead translates into throughput gains. The paper also reports that DeepRx can operate with lower pilot-power fractions than the classical iterative receiver, for example $i$7 instead of $i$8 in a single-user $i$9 QPSK setting (Rezaie et al., 25 Jun 2025).

Taken together, these extensions show that “DeepRx” no longer denotes only the original SIMO receiver. It denotes a receiver family organized around TTI-wide convolutional processing, pilot-aware input design, and direct soft-bit output, with task-specific front-end modifications for spatial multiplexing or superimposed pilots.

5. Interpretability, targeted adaptation, and failure discovery

A major later development is the use of influence functions for DeepRx adaptation. The 2025 paper on targeted fine-tuning studies DeepRx as a fully convolutional DNN-based wireless receiver and as a multi-label classifier whose training input/label pairing consists of received frequency-domain signals with transmitted bits as labels. The performance task is uncoded bit detection, evaluated by

$j$0

The paper defines targets as the evaluation samples with the largest relative BER gap to a genie-aided benchmark, restricted to a practical regime where $j$1 (Tuononen et al., 19 Sep 2025).

The classic influence of a training sample $j$2 on the loss at a target point $j$3 is written as

$j$4

with negative influence interpreted as beneficial and positive influence as harmful. The paper also introduces cross-loss influence, parameter-relative influence, loss-relative influence, and a self-influence correction called Newfluence. For DeepRx, the main practical result is that loss-relative influence is the most effective and consistent variant for fine-tuning (Tuononen et al., 19 Sep 2025).

The targeted fine-tuning procedure has four stages: train the baseline DeepRx model, pick target evaluation samples with the worst relative BER gap to genie, compute influence scores from every training sample to each target and select the most beneficial training samples, and briefly fine-tune DeepRx on those selected samples before reevaluating BER and BER-gap reduction. In the main experiments, the most beneficial training examples were selected from a top-50 pool and randomly ordered during fine-tuning. In single-target adaptation, first-order fine-tuning on beneficial influential samples reduces BER faster than random sample selection. On the worst target in the 800k regime, influence-guided fine-tuning achieved $j$5 relative BER-gap reduction versus $j$6 for random fine-tuning; in the 200k regime, the best average reduction was $j$7 for BCE + classic and $j$8 for BCE + $j$9-relative, while in the 800k regime the clearly best result was BCE + $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$0-relative at $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$1 (Tuononen et al., 19 Sep 2025).

The same study is equally clear about the boundaries of this method. The effect is short-lived: after around 15 fine-tuning steps, the advantage of influence-guided and random selection largely disappears because the original influence estimates become stale. Harmful samples do not yield a practical adaptation strategy: naive first-order gradient ascent on harmful examples caused catastrophic degradation, and the second-order influence-aligned update gave only negligible gains over first-order updates. Multi-target adaptation is much less effective, and influence-guided fine-tuning is consistently worse on non-target validation instances. The paper therefore concludes that influence-guided tuning is effective for single-target DeepRx adaptation but does not generalize well to multi-target scenarios (Tuononen et al., 19 Sep 2025).

A complementary line of work studies DeepRx not by adapting it, but by searching for operating conditions where it fails. The ATLAS framework treats DeepRx as the device under test and optimizes physically meaningful scenario variables—UE speed, channel delay spread, and channel noise power—inside a differentiable Sionna simulation, using the loss

$\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$2

For an under-trained DeepRx variant, ATLAS found 65 failure cases, 17 immediate failures, and 18 non-failing configurations across 100 episodes, and the paper states that all identified failure cases lie outside the trained channel parameter space. For more fully trained DeepRx variants, failures still exist but are fewer; they concentrate mostly in higher-SNR regions, leading the authors to suggest that DeepRx has strong denoising capability but may become comparatively weaker when noise is low and channel effects dominate (Belgiovine et al., 17 Aug 2025).

6. Efficiency, deployment drift, and operational constraints

Deployment-oriented DeepRx work has emphasized that receiver quality alone is not sufficient. The energy-efficiency case study analyzes the original fully convolutional ResNet-based SIMO version of DeepRx and separates total energy into memory and computation contributions,

$\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$3

The paper identifies the central residual blocks 5, 6, and 7 as the most energy-hungry blocks and uses that observation to compress the model by reducing the size of the central layers rather than shrinking the entire network uniformly. The main teacher model is 30 TFLOPs; the main student is 11 TFLOPs; teacher candidates range from 19 TFLOPs to 164 TFLOPs; and the best teacher for the 11 TFLOPs student is 30 TFLOPs, not the largest available model (Lbath et al., 14 Jul 2025).

The same paper applies knowledge distillation with

$\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$4

and reports best settings of $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$5 and $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$6 for the 11 TFLOPs student and 30 TFLOPs teacher. The distilled student exhibits a lower error floor than the same-size model trained from scratch, and the paper reports about a 4 dB gain at $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$7. The same study reports single-inference energy on Coral Edge TPU of 2 mJ/inference and notes that estimated and measured energy are broadly consistent in order of magnitude, while also stating that memory access causes deviations from simple FLOPs/Watt estimates (Lbath et al., 14 Jul 2025).

A more stringent operational critique appears in the 2026 calibration-drift benchmark using the public pretrained MathWorks DeepRx_2M model, a fully convolutional neural receiver with 1.23 million trainable parameters. In a 16-scenario 5G NR uplink PUSCH benchmark, the DeepRx-style neural receiver gains 1.0–2.0 dB over practical MMSE in only 3 scenarios, ties in 10 scenarios, underperforms by about 2 dB in QPSK, and fails silently at DMRS AddPos $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$8, an out-of-distribution but standards-compliant waveform configuration. In the strongest positive case, for delay spread 100 ns, practical MMSE requires 12.3 dB to reach 10% BLER while the neural receiver requires 10.3 dB; in the strongest negative case, DMRS AddPos $\mathbf{y}_{ij}\in\mathbb{C}^{N_r\times 1}$9, the neural receiver is marked fail while practical MMSE remains healthy (Elnashar, 24 May 2026).

That paper’s main systems proposal is detect-and-rollback: run the neural receiver and practical MMSE receiver in parallel, monitor either hard-bit disagreement or median-normalized LLR confidence on disagreeing bits, and switch back to the classical receiver when drift is detected. On the silent-failure DMRS AddPos $x_{ij}\in\mathbb{C}$0 scenario, the hard-bit disagreement detector recovers 10% BLER at 6.4 dB, matching the classical receiver within 0.02 dB. At 500 Hz Doppler, however, the same disagreement detector fails because the classical receiver is the one that collapses and the neural receiver is the one that succeeds. The paper therefore concludes that DeepRx should be treated as a monitored augmentation with a classical fallback, rather than as an unconditional replacement for practical MMSE (Elnashar, 24 May 2026).

In that sense, the history of DeepRx has evolved from architectural demonstration to broader receiver science. The original work established that a carefully designed fully convolutional neural receiver can outperform practical LMMSE-based processing in challenging OFDM settings (Honkala et al., 2020). Later work extended that design to MIMO detection (Korpi et al., 2020), spectrally efficient superimposed-DMRS uplink reception (Rezaie et al., 25 Jun 2025), targeted influence-based adaptation (Tuononen et al., 19 Sep 2025), failure-oriented test-and-measurement (Belgiovine et al., 17 Aug 2025), energy-aware compression (Lbath et al., 14 Jul 2025), and drift-aware rollback deployment strategies (Elnashar, 24 May 2026). This suggests that DeepRx is best understood not as a single network instance, but as a research program in learned radio reception centered on full-grid convolutional inference, soft-bit compatibility, and systematic comparison against classical receiver pipelines.