Abstract
Recent years have witnessed growing interest in the application of deep neural networks (DNNs) for receiver design, which can potentially be applied in complex environments without relying on knowledge of the channel model. However, the dynamic nature of communication channels often leads to rapid distribution shifts, which may require periodically retraining. This paper formulates a data-efficient two-stage training method that facilitates rapid online adaptation. Our training mechanism uses a predictive meta-learning scheme to train rapidly from data corresponding to both current and past channel realizations. Our method is applicable to any deep neural network (DNN)-based receiver, and does not require transmission of new pilot data for training. To illustrate the proposed approach, we study DNN-aided receivers that utilize an interpretable model-based architecture, and introduce a modular training strategy based on predictive meta-learning. We demonstrate our techniques in simulations on a synthetic linear channel, a synthetic non-linear channel, and a COST 2100 channel. Our results demonstrate that the proposed online training scheme allows receivers to outperform previous techniques based on self-supervision and joint-learning by a margin of up to 2.5 dB in coded bit error rate in rapidly-varying scenarios.
Introduction
In modern wireless communications, the physical environment—the "channel"—is rarely static. Factors such as user mobility and environmental changes cause frequent distribution shifts. While Deep Neural Network (DNN)-based receivers offer flexibility over traditional rigid mathematical models, they suffer from "data hunger." Typically, these models require massive amounts of pilot data to retrain whenever the channel shifts. In rapidly varying scenarios, the channel may change faster than the receiver can adapt, leading to significant performance degradation. This work addresses the challenge of enabling a deep learning receiver to adapt to dynamic channels quickly and with minimal data overhead.
Methodology
The proposed framework integrates meta-learning into the receiver's operation through a two-stage strategy designed for data efficiency and rapid adaptation.
Predictive Online Meta-Learning
Building on the Model-Agnostic Meta-Learning (MAML) framework, the algorithm treats the detection of the current block as a "support task" and the detection of the next block as a "query task." Instead of merely minimizing current error, it optimizes the model's initialization to capture temporal patterns of channel variations. This allows the model to adapt to the next channel realization with only a few gradient updates.
Self-Supervised Adaptation and Modularity
To minimize bandwidth overhead, the receiver utilizes self-supervised adaptation. If a data block passes a reliability check (such as a CRC), the decoded bits are re-encoded and used as "pseudo-labels" for on-the-fly retraining. Furthermore, for hybrid architectures like DeepSIC—which combine model-based Soft Interference Cancellation with DNNs—the authors propose modular training. By updating only specific "dynamic" modules (e.g., those corresponding to a moving user) while keeping stationary modules fixed, the system achieves faster convergence and reduces the number of parameters to be learned.
Experimental Evaluation
The approach was evaluated across three distinct channel profiles:
- Synthetic Linear Gaussian Channel: A baseline with oscillating frequency taps.
- Synthetic Non-Linear Channel: Incorporates a
tanhtransformation to simulate hardware-induced non-linearities like power amplifier distortion. - COST 2100 Channel: An industry-standard geometry-based stochastic model for realistic 5 GHz indoor environments.
Key Results and Discussion
The proposed meta-learning scheme demonstrated significant gains, outperforming standard self-supervision and joint-learning by up to 2.5 dB in coded bit error rate (BER) in rapidly varying scenarios. These gains were consistent even in non-linear channels where traditional model-based receivers typically fail.
While meta-learning introduces per-block computational overhead, it reduces the total number of gradient steps required for convergence compared to standard online training. In MIMO setups with mixed static and mobile users, modular training provided an additional 0.3 dB gain and faster adaptation than updating the entire network.
Conclusion
By using meta-learning as a predictor for channel evolution, this framework allows receivers to "skate to where the puck is going to be." While the effectiveness relies on underlying structure in channel variations, the hybrid model-based approach provides a necessary "skeleton" of domain knowledge that makes meta-learning efficient. Future work will focus on automatically detecting mobility to trigger modular updates and implementing these algorithms on AI hardware accelerators to meet the sub-millisecond latency requirements of 5G/6G networks.
Summary updated