MetaSSD: Meta-Learned Self-Supervised Detection

Moon Jeong Park, Jungseul Ok, Yo-Seb Jeon, Dongwoo Kim

Abstract

In modern wireless communications, the demand for high-speed data processing often conflicts with the computational and data requirements of deep learning models. Traditional symbol detectors, while mathematically optimal, are either computationally prohibitive or highly sensitive to channel estimation errors. This paper discusses MetaSSD (Meta-Learned Self-Supervised Detection), a novel framework that integrates Model-Agnostic Meta-Learning (MAML) with self-supervised reconstruction. By establishing a robust parameter initialization, MetaSSD adapts to rapidly changing channel conditions using significantly fewer pilot symbols than existing deep learning solutions, maintaining high accuracy even in the presence of noisy channel state information.

Introduction

Deep learning has emerged as a promising alternative to traditional signal processing in wireless systems. However, two primary hurdles remain: the "data-hungry" nature of neural networks and the volatility of wireless channels. Standard deep learning detectors often require thousands of pilot symbols—known data used for training—to adapt to each new channel state. This overhead consumes valuable bandwidth and introduces latency.

MetaSSD addresses these challenges by leveraging meta-learning to find a "universal" starting point for model weights. This allows the detector to achieve peak performance with 50 to 100 times less training data than previous state-of-the-art models like ViterbiNet. Furthermore, by incorporating a self-supervised reconstruction loss, MetaSSD ensures that its predictions remain physically consistent with the received signal.

The MetaSSD Framework

1. Meta-Learning via MAML

Rather than training a model for a specific static environment, MetaSSD employs Model-Agnostic Meta-Learning (MAML). During the meta-training phase, the model is exposed to a diverse range of simulated channel realizations. The goal is not to master a single channel but to learn a weight initialization that is highly sensitive to the underlying physics of symbol detection. Consequently, when the model encounters a new, unseen channel, it can reach an optimal state in as few as four gradient update steps.

2. Self-Supervised Reconstruction Loss

Supervised learning in wireless detection is limited by the number of available pilot symbols. MetaSSD augments this by adding a self-supervised reconstruction loss. The model takes its own symbol predictions, passes them through a known (or estimated) channel model, and compares the result to the actual received signal. This feedback loop allows the model to learn from the entire data block, effectively using the received data itself as a training signal to regularize the learning process and prevent overfitting to small pilot sets.

3. Learnable Temperature Scaling

To account for the inevitable inaccuracies in channel estimation, the authors introduced a learnable temperature parameter ($\tau$) within the softmax layer. This parameter allows the network to adjust the "sharpness" of its probability distributions. In high-noise environments or when channel information is unreliable, the model can automatically smooth its predictions, providing a layer of robustness that traditional algorithms lack.

Experimental Methodology

The performance of MetaSSD was evaluated using synthetic Inter-Symbol Interference (ISI) channels characterized by frequency-selective Rayleigh fading. The simulation parameters included:

  • Modulation: Binary Phase Shift Keying (BPSK).
  • Scale: Meta-training across 10,500 channel realizations.
  • Testing: 500 datasets per SNR level (0 to 15 dB).
  • Efficiency Benchmark: MetaSSD was restricted to only 100 training symbols, whereas baseline deep learning models typically require 5,000 to 10,000 symbols.

Key Results and Discussion

Performance vs. Traditional Baselines

MetaSSD consistently outperformed the standard OFDM-MMSE (Minimum Mean Square Error) detector across the entire SNR spectrum. More significantly, MetaSSD demonstrated superior resilience compared to the BCJR algorithm—the mathematically optimal detector for ISI channels—when channel estimation was noisy. For SNR levels $\ge$ 8 dB, MetaSSD achieved a lower Symbol Error Rate (SER) than BCJR, proving that deep learning can surpass "optimal" math-heavy models by better handling real-world estimation errors.

Data Efficiency and Robustness

The integration of MAML and self-supervision proved critical. Ablation studies indicated that removing either component resulted in a significant degradation of performance. The synergy between a well-primed initialization and a physical reconstruction constraint allows MetaSSD to bridge the gap between perfect and noisy channel knowledge, making it a viable candidate for low-latency, real-time wireless applications.

Conclusion and Future Directions

MetaSSD represents a significant step toward making deep learning practical for wireless physical layer tasks. By reducing the pilot symbol requirement by two orders of magnitude, it preserves bandwidth while maintaining high detection accuracy. Future research will focus on extending this framework to higher-order modulation schemes (such as 64-QAM) and validating the model against real-world over-the-air (OTA) data to ensure robustness against non-synthetic channel impairments.

Summary updated

BibTeX
@inproceedings{park2022metassd,
  author = {Moon Jeong Park and Jungseul Ok and Yo-Seb Jeon and Dongwoo Kim},
  title = {{MetaSSD: Meta-Learned Self-Supervised Detection}},
  booktitle = {ISIT},
  year = {2022},
  doi = {10.1109/isit50566.2022.9834613},
}
Figure