Deep Learning-Aided SIC

• Deep Learning-Aided SIC is a method that fuses data-driven neural networks with traditional successive interference cancellation to overcome model mismatches and reduce pilot overhead.
• It employs architectures like deep unfolding, blockwise stacking, and graph neural networks to enhance detection, equalization, estimation, and semantic decoding in multiuser systems.
• DL-SIC achieves superior performance metrics in scenarios such as NOMA and MIMO, demonstrating lower SER, reduced device complexity, and rapid adaptation via meta-learning.

Deep learning-aided successive interference cancellation (DL-SIC) refers to a family of detection, equalization, estimation, and semantic decoding schemes in which core interference cancellation steps are replaced or augmented by data-driven neural networks, enabling greater flexibility, improved robustness to channel/model mismatch, and significant gains in sample- and pilot-efficiency. DL-SIC methodologies have been deployed across diverse wireless contexts including non-orthogonal multiple access (NOMA), MIMO detection, soft interference cancellation, nonlinear equalization, joint pilot/channel recovery, and semantic multiuser communication. These approaches integrate deep neural networks (DNNs)—often through deep unfolding, meta-learning, or graph-based parameter sharing—directly into the recursive structure of conventional SIC, yielding end-to-end trainable architectures that match or exceed model-based performance, frequently with lower on-device complexity and improved adaptability.

1. Key Architectures: From Model-Based SIC to Deep Learning Integration

Classical SIC recursively decodes or estimates users' symbols or channel contributions, removing previously decoded signals at each stage. Its core bottlenecks are model dependence, sensitivity to channel state information (CSI) uncertainty, limitations under severe nonlinearities, and the need for perfect channel estimates or extensive pilot resources.

DL-SIC replaces deterministic model-based blocks—linear symbol estimation, hard slicing, covariance computation, and Gaussian likelihood calculation—with DNNs or learned modules. Two principal paradigms recur:

• Deep Unfolding: The iterative update equations of soft SIC are mapped onto a feedforward or recurrent neural network, with each SIC (or detection/equalization) stage corresponding to a network layer or cell, as in DeepSIC (Shlezinger et al., 2020), SICNN (Baumgartner et al., 2023), and GAN-SIC (Nguyen et al., 2022).
• Blockwise Feedforward Stacking: DNN blocks are sequenced such that each block handles one user or stream, taking as input the received observations and soft outputs from previous blocks (SICNet (2207.14468), meta-SICNet (Issa et al., 2023), recurSIC (Fesl et al., 23 Jan 2026)).
• Graph Neural Networks: Parameter sharing and message passing over multiuser or multi-path graphs (GNNSIC (Yi et al., 13 Feb 2026)) enable scale- and sample-efficient generalization, crucial for large-scale deployments.

These architectures often utilize softmax layers for symbol probabilities, ReLU or batch-norm for hidden nonlinearity, cross-entropy for loss, and Adam or SGD variants for optimization.

2. Algorithmic and Mathematical Formulation

DL-SIC systems obey the superposition/broadcast/additive channel models of their communication scenario:

• Multiuser MIMO Uplink/Downlink:
  • Signal model: $y = Hs + n$, with $s\in\mathcal{S}^K$, $n\sim\mathcal{CN}(0,\sigma^2 I)$.
  • Classical SIC: hard ML slicing and recursive subtraction, or soft iterative updates based on posterior/likelihood computation, relying on known or estimated $H$.
  • DL-SIC: DNN-based blocks operate on concatenations of $y$, soft symbol vectors from past stages, channel feature vectors, and SNR embeddings (2207.14468, Issa et al., 2023, Fesl et al., 23 Jan 2026). Output is the posterior $$\mathbf{p}_l = \left[\hat{p}(x_l=\gamma_1), \ldots, \hat{p}(x_l=\gamma_M)\right]$$ for constellation symbols.
• Meta-Learning for Few-Pilot SIC:
  • Meta-SICNet employs a MAML-style update, where meta-parameters $\theta$ are optimized across tasks (devices), and rapid adaptation to new devices occurs with as few as 2-4 pilots via gradient steps on the local task loss $L_{\mathcal{D}_T}(\theta)$ (Issa et al., 2023).
• Soft-Output (LLR) Computation:
  • In recursive, multi-path SIC, candidate symbol lists are generated per stage. LLR values are produced using the max-log rule over minimum metric differences among competing bit hypotheses (Fesl et al., 23 Jan 2026).
• GNN-based SIC:
  • GNNSIC defines user nodes and their interference as edges; per-iteration DNN modules (embedding, messaging, aggregation, update) share parameters across all users and iterations, greatly reducing parameter complexity and improving generalization (Yi et al., 13 Feb 2026).
• Nonlinear/Non-Gaussian Channels:
  • Deep/unfolded RNNs approximate the forward-backward algorithm (FBA) or BCJR equalizer in each SIC stage, with input windows encompassing received sample histories and previously decoded symbols (Plabst et al., 2024).

3. Performance and Complexity Analysis

DL-SIC consistently improves core metrics relative to classical SIC and nonunfolded or non-shared deep detectors:

System / Scenario	Performance Metric	DL-SIC Gain/Result	Reference
NOMA (few-pilot)	SER vs. pilot count, SNR	meta-SICNet: SER$\approx 10^{-2}$ w/ $P=2$ pilots, $s\in\mathcal{S}^K$0 lower SER than classical SIC (imperfect CSI)	(Issa et al., 2023)
MIMO (16QAM/64QAM)	Hard, soft BER, complexity	recurSIC: $s\in\mathcal{S}^K$1dB gain over MMSE-SIC; near-ML (<0.5dB); $s\in\mathcal{S}^K$2 parameters	(Fesl et al., 23 Jan 2026)
Linear/Nonlinear MIMO	SER, CSI uncertainty	DeepSIC, GNNSIC more robust than analytic SIC; GNNSIC achieves low SER with $s\in\mathcal{S}^K$3\% parameters of DeepSIC	(Yi et al., 13 Feb 2026)
Full-duplex analog SIC	Cancellation depth, BER	DL+STE achieves $s\in\mathcal{S}^K$4 dB cancellation, robust under ADC saturation	(Knaepper et al., 11 Mar 2025)
Frequency-domain equalization	BER, computation	SICNNv1 beats all model- and unfolded-NN methods for SC-FDE, $s\in\mathcal{S}^K$5 weights; reduced parameter versions preserve SNR slope	(Baumgartner et al., 2023)
Semantic MAC (multiuser)	BLEU, semantic similarity	Semantic SIC achieves high BLEU and semantic match at $s\in\mathcal{S}^K$6 dB lower SNR vs. bit-centric SIC, rapid adaptation with pretraining/partial retraining	(Li et al., 19 Jan 2025)

DL-SIC's parameter sharing, as in GNNSIC, eliminates the exponential penalty in Rademacher complexity associated with depth or SIC stages, promoting sample-efficient training (Yi et al., 13 Feb 2026). In practical terms, $s\in\mathcal{S}^K$7 samples suffice for reliable operation in moderate-size MIMO, versus $s\in\mathcal{S}^K$8–$s\in\mathcal{S}^K$9 for unshared unfolded architectures.

4. Adaptation, Meta-Learning, and Online/Hybrid Training

Adaptability to channel variations and device heterogeneity is a hallmark feature of DL-SIC methods:

• Meta-SICNet leverages MAML to establish a meta-initialization over diverse channel/device families, then enables per-device adaptation with as few as 2-4 pilots and minimal online computation (e.g., $n\sim\mathcal{CN}(0,\sigma^2 I)$0 ms/episode) (Issa et al., 2023).
• SICNet (pilotless online retraining): FEC decoding-based label recovery supports SGD adaptation under block fading without new pilots, tracking fading at negligible pilot/data overhead (2207.14468).
• GAN-SIC: Integrates a conditional GAN for synthetic data distribution tracking, enabling self-supervised, label-free online retraining of the SIC detector in dynamic MIMO channels. This approach achieves significantly lower SER on rapidly varying or non-Gaussian channels compared to retrained DeepSIC or classical SIC (Nguyen et al., 2022).
• Semantic SIC: Partial retraining strategies with frozen pre-trained modules facilitate adaptation when new users join, preserving low complexity and rapid convergence (Li et al., 19 Jan 2025).

5. Extensions to Nonlinear, Bandlimited, and Semantic Channels

DL-SIC generalizes naturally to scenarios beyond linear Gaussian models:

• Nonlinear bandlimited channels: Periodic, time-varying bidirectional RNNs in the-SIC stages approximate forward-backward marginalization, attaining within 1 dB of joint detection and decoding upper bounds with orders-of-magnitude reduced complexity (e.g., $n\sim\mathcal{CN}(0,\sigma^2 I)$1 multiplications vs. exponential in trellis memory for BCJR) (Plabst et al., 2024).
• Analog self-interference cancellation in full-duplex radios: MLP-based forward Hammerstein models, trained via MSE on ADC-residual outputs, achieve robust SIC even when omitting ADC layer gradients (straight-through estimation); design guidelines recommend STE for low complexity/adaptive applications (Knaepper et al., 11 Mar 2025).
• Semantic communication over MAC: Deep transformer and autoencoder stacking realizes a semantic SIC at the feature-embedding level. Performance is substantially boosted by joint/partial retraining, integrated cross-user feature aggregation, and side-information, producing SNR and accuracy improvements unattainable with symbol-level SIC (Li et al., 19 Jan 2025).

6. Training, Regularization, and Sample Efficiency

DL-SIC architectures are universally trained end-to-end with cross-entropy (symbol or soft-output) or MSE (for regression tasks) as the main loss functions:

• Loss-weighting strategies (e.g., "combined" vs. "local" in SICNet; SNR-gridded training in SICNN) enhance convergence, regularization, or robustness to order/power variations (2207.14468, Baumgartner et al., 2023).
• Parameter sharing across users/stages/iterations (GNNSIC, reduced SICNNv1/v2) dramatically cuts the training sample requirement for a fixed error bound (Baumgartner et al., 2023, Yi et al., 13 Feb 2026).
• Regularization techniques (batch norm, weight decay, early stopping) are occasionally used but often not essential given the modest model sizes and robustness imbued by architectural design.
• Self-supervised adaptation (block fading online adaptation, GAN sample re-generation) enables continual retraining with no explicit ground-truth labeling (2207.14468, Nguyen et al., 2022).

7. Limitations, Open Challenges, and Prospective Directions

While DL-SIC methodologies address many limitations of model-based SIC and classical deep detection, several challenges and frontiers remain (Issa et al., 2023, 2207.14468, Yi et al., 13 Feb 2026):

• Task distribution coverage: Meta-learning or domain-adaptive methods require carefully curated/representative task families; performance degrades if novel scenarios diverge sharply from training distribution.
• Complexity in massive user scenarios: Unfolded architectures without aggressive parameter sharing or GNN-based design scale poorly; advances in graph-based DL-SIC mitigate this but further theoretical and practical scaling results are needed.
• Extension to higher-order modulations, large antenna arrays, and real-time constraints requires ongoing work on lightweight or quantized architectures (e.g., pruning, shared-weight CNNs, parameter tying).
• Combined channel-estimation and detection: Integration of SIC-aided neural receivers with joint pilot design yields lower MSE/channel estimation than LMMSE methods, but unified frameworks for joint estimation/detection/meta-adaptation remain an active area (Chun et al., 2018, Issa et al., 2023).
• Self-interference and nonlinearities: While full-duplex and nonlinear equalization applications of DL-SIC are established, further translation to multi-carrier, MIMO, and wideband radio remains open.

• (Issa et al., 2023) Meta-Learning Based Few Pilots Demodulation and Interference Cancellation For NOMA Uplink
• (Fesl et al., 23 Jan 2026) Learning Successive Interference Cancellation for Low-Complexity Soft-Output MIMO Detection
• (2207.14468) Deep Learning Based Successive Interference Cancellation for the Non-Orthogonal Downlink
• (Yi et al., 13 Feb 2026) Data-Driven Deep MIMO Detection: Network Architectures and Generalization Analysis
• (Knaepper et al., 11 Mar 2025) On Digital Optimization of Analog Self-Interference Cancellation for Full-Duplex Wireless Systems
• (Baumgartner et al., 2023) SICNN: Soft Interference Cancellation Inspired Neural Network Equalizers
• (Shlezinger et al., 2020) DeepSIC: Deep Soft Interference Cancellation for Multiuser MIMO Detection
• (Nguyen et al., 2022) Interference Cancellation GAN Framework for Dynamic Channels
• (Plabst et al., 2024) Neural Network-Based Successive Interference Cancellation for Non-Linear Bandlimited Channels
• (Chun et al., 2018) Deep Learning Based Joint Pilot Design and Channel Estimation for Multiuser MIMO Channels
• (Li et al., 19 Jan 2025) A Semantic Approach to Successive Interference Cancellation for Multiple Access Networks