Learning to Demodulate from Few Pilots via Offline and Online Meta-Learning

Sangwoo Park, Hyeryung Jang, Osvaldo Simeone, Joonhyuk Kang

Abstract

This paper considers an Internet-of-Things (IoT) scenario in which devices sporadically transmit short packets with few pilot symbols over a fading channel. Devices are characterized by unique transmission non-idealities, such as I/Q imbalance. The number of pilots is generally insufficient to obtain an accurate estimate of the end-to-end channel, which includes the effects of fading and of the transmission-side distortion. This paper proposes to tackle this problem by using meta-learning. Accordingly, pilots from previous IoT transmissions are used as meta-training data in order to train a demodulator that is able to quickly adapt to new end-to-end channel conditions from few pilots. Various state-of-the-art meta-learning schemes are adapted to the problem at hand and evaluated, including Model-Agnostic Meta-Learning (MAML), First-Order MAML (FOMAML), REPTILE, and fast Context Adaptation VIA meta-learning (CAVIA). Both offline and online solutions are developed. In the latter case, an integrated online meta-learning and adaptive pilot number selection scheme is proposed. Numerical results validate the advantages of meta-learning as compared to training schemes that either do not leverage prior transmissions or apply a standard joint learning algorithms on previously received data.

Introduction

In the burgeoning Internet of Things (IoT) landscape, devices often transmit short data packets to conserve energy and bandwidth. To facilitate demodulation, transmitters include "pilot" symbols that allow the receiver to estimate Channel State Information (CSI). However, the extreme overhead constraints in IoT mean that only a handful of pilots are available. Furthermore, low-cost IoT hardware introduces significant "non-idealities," such as I/Q imbalance—mismatches between the real and imaginary signal components—that create complex, non-linear distortions.

System model showing meta-training devices and a meta-test device communicating with a base station.

Traditional model-based estimation techniques, such as Minimum Mean Square Error (MMSE), often struggle in these regimes because the limited number of pilots is insufficient to characterize both the stochastic fading channel and device-specific hardware distortions. While deep learning offers a potential solution, training a neural network from scratch for every transmission is infeasible with only 2 or 4 pilot symbols.

Methodology

The authors propose a Meta-Learning framework to address this data scarcity. By "learning to learn" from a history of previous transmissions, a base station can develop an inductive bias that allows it to rapidly specialize a demodulator for a new device using minimal pilot data.

Meta-Learning Algorithms

Several state-of-the-art algorithms were adapted for the demodulation task:

  • MAML (Model-Agnostic Meta-Learning): Optimizes for a weight initialization that enables high accuracy after just one or two gradient descent steps on new pilots.
  • FOMAML & REPTILE: Computationally efficient first-order approximations of MAML.
  • CAVIA (Context Adaptation VIA meta-learning): Instead of shifting all network weights, CAVIA adapts a small "context" vector. This approach is particularly robust against overfitting when pilot data is extremely limited.

Offline and Online Frameworks

The study explores two operational modes. In the offline setting, the model is trained on a fixed database of past transmissions. In the online setting, the base station continuously updates its meta-knowledge as new devices connect. A key innovation in the online framework is an adaptive pilot selection scheme. Here, the base station evaluates the reliability of its current demodulator and dynamically instructs the next device on the required number of pilots, optimizing the trade-off between overhead and accuracy.

Flowchart of the integrated online meta-learning and adaptive pilot number selection scheme.

Experimental Setup and Results

The framework was evaluated using simulated datasets, including Binary Fading (4-PAM) and Rayleigh Fading with I/Q imbalance (16-QAM). Hardware distortions were modeled using amplitude and phase imbalance factors drawn from Beta distributions.

Key Findings: * Performance Gains: In 16-QAM scenarios, meta-learning methods significantly outperformed traditional MMSE channel estimation combined with ML demodulation, especially when the number of pilots was below 16. * Robustness of CAVIA: CAVIA was found to be superior to MAML when pilots were extremely scarce (e.g., 4 pilots for 16-QAM), as its limited parameter adaptation prevents the model from "hallucinating" patterns in noise. * Overhead Reduction: The adaptive online scheme successfully maintained target symbol error rates while using fewer pilots on average than non-adaptive methods, demonstrating a clear path to reducing IoT network overhead.

Conclusion and Future Work

Meta-learning provides a powerful mechanism for receivers to handle the hardware non-idealities and short-packet constraints of IoT. By pre-configuring the receiver with a "neutral" yet flexible starting point, the system achieves near-optimal demodulation with minimal overhead.

Future research may explore the integration of meta-learning with Compressive Sensing (CS) to address more complex scenarios, such as multi-antenna (MIMO) or frequency-selective channels. Additionally, addressing the requirement for a downlink control channel in adaptive schemes remains a practical challenge for the simplest uplink-only IoT protocols.

Summary updated

Abstract

This paper considers an Internet-of-Things (IoT) scenario in which devices sporadically transmit short packets with few pilot symbols over a fading channel. Devices are characterized by unique transmission non-idealities, such as I/Q imbalance. The number of pilots is generally insufficient to obtain an accurate estimate of the end-to-end channel, which includes the effects of fading and of the transmission-side distortion. This paper proposes to tackle this problem by using meta-learning. Accordingly, pilots from previous IoT transmissions are used as meta-training data in order to train a demodulator that is able to quickly adapt to new end-to-end channel conditions from few pilots. Various state-of-the-art meta-learning schemes are adapted to the problem at hand and evaluated, including Model-Agnostic Meta-Learning (MAML), First-Order MAML (FOMAML), REPTILE, and fast Context Adaptation VIA meta-learning (CAVIA). Both offline and online solutions are developed. In the latter case, an integrated online meta-learning and adaptive pilot number selection scheme is proposed. Numerical results validate the advantages of meta-learning as compared to training schemes that either do not leverage prior transmissions or apply a standard joint learning algorithms on previously received data.
BibTeX
@article{park2021learning,
  author = {Sangwoo Park and Hyeryung Jang and Osvaldo Simeone and Joonhyuk Kang},
  title = {{Learning to Demodulate from Few Pilots via Offline and Online Meta-Learning}},
  journal = {IEEE Trans. Signal Process.},
  year = {2021},
  url = {https://openreview.net/forum?id=cKxu4GFQU9},
}
Figure