Source-linked AI summary
Learning to Detect
Neev Samuel, Tzvi Diskin, Ami Wiesel
TL;DR
The paper investigates whether deep neural networks can provide computationally efficient MIMO detection near optimal accuracy. It introduces two learning-based detector architectures and reports strong performance across channel models and constellations, including soft outputs and single-training detection over multiple channels.
Problem
MIMO detection involves complex-valued signals and varied digital constellations, motivating a unified machine-learning framework for detection.
Method
The paper formulates detection as minimizing expected loss over the MIMO model distribution and introduces two deep learning networks for MIMO detection.
Results
The networks achieve near optimal detection performance with low computational complexity across different channel models and digital constellations, while also providing accurate soft outputs and detecting multiple channels after a single training.
Takeaways & Limitations
Deep neural networks can serve as MIMO detectors with promising accuracy, flexible computational complexity, soft posterior outputs, and one network trained across multiple channel realizations.
Abstract
from arXiv · showhide
In this paper we consider Multiple-Input-Multiple-Output (MIMO) detection using deep neural networks. We introduce two different deep architectures: a standard fully connected multi-layer network, and a Detection Network (DetNet) which is specifically designed for the task. The structure of DetNet is obtained by unfolding the iterations of a projected gradient descent algorithm into a network. We compare the accuracy and runtime complexity of the purposed approaches and achieve state-of-the-art performance while maintaining low computational requirements. Furthermore, we manage to train a single network to detect over an entire distribution of channels. Finally, we consider detection with soft outputs and show that the networks can easily be modified to produce soft decisions.
I. INTRODUCTION
MIMO detection offers communication benefits but is computationally difficult because optimal detection can require exponential or exhaustive search. This paper introduces deep-learning detectors designed to approach optimal accuracy with low complexity across channels and constellations, including soft outputs and single-training multi-channel detection.
- MIMO systems improve communication performance by exploiting multiple dimensions, but detection is NP-Complete and motivates sub-optimal polynomial-complexity solutions.
- Maximum-likelihood detection is optimal for joint symbol error probability but has exponential runtime complexity, making it impractical for large real-time systems.
- The paper introduces two deep-learning networks for MIMO detection and evaluates them across channel models and digital constellations.
- The networks can provide soft outputs with accuracy comparable to an M-Best sphere decoder and low computational complexity.
- DetNet can handle multiple channel models with a single training phase, unlike prior cited approaches developed for a single fixed channel.
D. Notation
The paper establishes notation for distributions, vectors, matrices, indexing, Gaussian matrices, nonlinearities, and complex-valued quantities, then states the standard linear MIMO model and a noise-variance assumption.
- Normal and uniform distributions are denoted using their parameters, while uppercase and lowercase bold symbols represent matrices and vectors.
- The superscript T denotes transpose, and x_i denotes the i’th element of vector x.
- An i.i.d. Gaussian matrix has entries sampled independently from N(0,1), and the rectified linear unit is ρ(x) = max{0, x}.
- For complex matrices and vectors, ℜ(·) and ℑ(·) denote real and imaginary parts, respectively.
- The standard linear MIMO model uses received vector y, channel matrix H, transmitted symbol vector x, and Gaussian noise w.
- The detectors do not assume knowledge of noise variance σ^2, unlike MMSE and AMP decoders that exploit this parameter.
B. Reparameterization
The paper re-parameterizes complex-valued MIMO detection into real-valued vectors and one-hot symbol representations to use standard machine-learning tools in a unified framework.
- Complex-valued signals are represented using real-valued received vectors y ∈ R^2N, channel matrices H ∈ R^2N×2K, and symbol vectors x ∈ S^2K.
- Each constellation symbol is associated with a unit vector through a one-hot mapping, and the mapping is denoted s = f_oh(u).
- For approximate inputs that are not unit vectors, the one-hot-to-symbol function is defined more generally.
- A vector of 2K symbols is handled by stacking the one-hot representations of its elements, mapping x_oh ∈ {0, 1}^{|S|·2K} to x ∈ S^2K.
C. Learning to detect
The paper formulates detection as expected-loss minimization over a MIMO distribution and compares a general fully connected network with architectures designed for channel-dependent detection. FullyCon is simple and effective for fixed channels but inadequate for varying channels, motivating a more expressive design.
- C. Learning to detect: A detector is modeled as a parameterized architecture x̂_oh(H, y; θ), with learning choosing θ to minimize loss over the MIMO model distribution.
- C. Learning to detect: The analysis assumes perfect channel state information, distinguishing fixed channels from varying channels drawn from a known continuous distribution.
- FullyCon: FullyCon is a standard fully connected multilayer network in which each layer’s output feeds the next layer.
- FullyCon: FullyCon uses only y as direct input and represents estimates through a final mapping from the network output.
- FullyCon: The network is trained with an l2 distance between the estimated and true signals.
- FullyCon: FullyCon achieves almost optimal accuracy with low complexity for fixed channels but performs poorly for varying channels because it cannot capture changing-channel dependencies.
B. DetNet
DetNet is designed specifically for MIMO detection by unfolding projected-gradient iterations into a trainable deep architecture with learned step sizes and layerwise training support.
- DetNet is a deep architecture designed specifically for MIMO detection.
- Its construction uses H^T y and H^T Hx_k as key ingredients rather than processing y directly.The architecture uses a compressed sufficient statistic and gradient-like updates.
- Each iteration combines the current estimate with gradient terms and applies a nonlinear projection after lifting to a higher dimension.
- The gradient step sizes δ_k are learned parameters optimized during training.
- The final estimate is x̂_L, with W1_k defined as tall and skinny matrices for lifting and nonlinearities.
- DetNet uses a loss incorporating outputs from all layers to address training difficulties such as vanishing gradients and activation saturation.
- A residual feature further enhances DetNet by making each layer’s output a weighted average with the previous layer’s output.
IV. SOFT DECISION OUTPUT
The paper extends MIMO detection to soft outputs because iterative decoders require posterior distributions, while direct posterior computation is exponentially complex. It shows that neural networks can approximate these posteriors through one-hot outputs trained with standard l2 loss.
- Soft-output MIMO detectors provide posterior probabilities for each unknown and possible constellation symbol to support iterative decoding.
- Direct posterior computation is optimal but has complexity exponential in signal and constellation size, motivating approximate posterior-output networks.
- With sufficient expressiveness and global optimization, the one-hot output x̂_oh provides exact posterior probabilities.
V. NUMERICAL RESULTS
The numerical-results section evaluates the proposed networks on fixed and varying channel settings, including random channels for varying-channel and soft-output experiments.
- The experiments compare the proposed networks with competing methods on accuracy and computational complexity.
- Fully connected results are reported for the 0.55-Toeplitz channel.
- Varying-channel and soft-output results use random channels whose entries are independently sampled from N(0, 1).
A. Implementation details
The implementation uses Adam and TensorFlow, trains on independently generated samples across uniformly distributed SNRs, and evaluates named neural and classical detector variants.
- Both networks are trained with the Adam optimizer and implemented using TensorFlow.
- Training samples are independently generated according to the signal, channel, and noise statistics.
- A DetNet layer’s output is passed as the input to the next layer across L stacked layers.
- The training noise variance is randomized so that SNR is uniformly distributed over U(SNR_min, SNR_max).
- The implementation names the basic fully connected architecture FullyCon and the task-specific architecture DetNet.
- Hard-decision evaluations include ZF, AMP, SDR, and sphere decoding baselines.
- Soft-output evaluations compare against M-Best sphere decoding with M=5 and M=7 candidates.
C. Accuracy results
DetNet performs near the strongest traditional detectors across fixed and varying-channel settings, while soft-output accuracy improves with additional layers and varies by constellation.
- Fixed channel: DetNet and FullyCon achieve accuracy comparable to SDR and AMP in the fixed-channel scenario.In this setting, both networks perform comparably to most competitors except SD.
- Fixed channel: AMP did not converge in the fixed-channel evaluation, which is described as a difficult setting.Figure 3 reports accuracy over a range of SNR values for the fixed-channel model.
- Varying channel: DetNet achieves accuracy comparable to SDR and AMP and near SD for varying 30 × 60 BPSK channels.Figure 4 evaluates BER across varying real-valued channels of size 30 × 60.
- Varying channel: DetNet achieves accuracy comparable to SDR and AMP and near SD for varying 20 × 30 QPSK channels.Figure 5 evaluates BER across varying complex channels of size 20 × 30.
- Soft outputs: For 8-PSK soft outputs, DetNet is comparable to M-Best algorithms only in the high-SNR region.The 16-QAM soft-output comparison reports comparable performance to M-Best Sphere decoding.
D. Computational Resources
The paper evaluates runtime across detector types, channel settings, constellation sizes, batch sizes, and hard- versus soft-output detection. DetNet generally benefits from batching, while its relative advantage varies with constellation and comparison method.
- Experimental setup: Runtime comparisons used a common Python 2.7 and NumPy environment, while reporting several batch sizes because batching gains depend on platform.The evaluation included batch size one and accounted for SNR-dependent sphere-decoding runtime.
- Fixed-channel hard decisions: FullyCon is faster than all other detectors in the fixed-channel hard-decision case, even without batching.DetNet is slightly faster than traditional detectors without batching and improves significantly with batches.
- Variable-channel hard decisions: In the variable-channel setting, DetNet’s relative advantage over AMP and SDR is smaller for 8-PSK and 16-QAM than for BPSK.For 16-QAM, AMP was slightly faster without batching, while DetNet’s relative performance versus sphere decoding improved.
- Soft outputs: For soft outputs, DetNet is comparable to M-Best sphere decoders for BPSK without batching and slightly faster for 16-QAM and 8-PSK.Batching significantly improves DetNet’s soft-output runtime in the BPSK case.
2) Accuracy-Complexity Trade-Off:
DetNet exposes a runtime accuracy-complexity trade-off because every layer produces an estimate and the loss trains all layer outputs. Earlier layers can be selected for faster detection, although later layers are usually more accurate.
- 2) Accuracy-Complexity Trade-Off:: Each DetNet layer outputs an estimated signal, and the loss function optimizes all layer outputs.The final layer is usually used because it is the most accurate.
- 2) Accuracy-Complexity Trade-Off:: Selecting an earlier layer enables faster detection by reducing the computation used at runtime.The paper presents accuracy as a function of the chosen output layer.
- 2) Accuracy-Complexity Trade-Off:: The networks provide promising accuracy with low and flexible computational complexity across digital constellations and can produce accurate soft posterior outputs.One network can detect over multiple channel realizations after a single training process.