Meta-Learning to Communicate: Fast End-to-End Training for Fading Channels

Sangwoo Park, Osvaldo Simeone, Joonhyuk Kang

Abstract

When a channel model is available, learning how to communicate on fading noisy channels can be formulated as the unsupervised training of an autoencoder consisting of the cascade of encoder, channel, and decoder. An important limitation of the approach is that training should be generally carried out from scratch for each new channel. To cope with this problem, prior works considered joint training over multiple channels with the aim of finding a single pair of encoder and decoder that works well on a class of channels. As a result, joint training ideally mimics the operation of non-coherent transmission schemes. In this paper, we propose to obviate the limitations of joint training via meta-learning: Rather than training a common model for all channels, meta-learning finds a common initialization vector that enables fast training on any channel. The approach is validated via numerical results, demonstrating significant training speed-ups, with effective encoders and decoders obtained with as little as one iteration of Stochastic Gradient Descent.

Introduction

In modern wireless research, the communication link is increasingly modeled as an end-to-end autoencoder. This architecture comprises a transmitter (encoder) that maps messages to signals, a physical channel that introduces noise and fading, and a receiver (decoder) that reconstructs the original message. While this approach is powerful, it faces a significant hurdle in dynamic environments: fading channels. When channel characteristics change due to mobility or environmental shifts, the autoencoder typically requires retraining from scratch, which is computationally prohibitive for real-time applications.

Existing solutions often fall into two categories:

  1. Per-channel Training: Retraining for every new condition, which is too slow.
  2. Joint Training: Training a single robust model across a variety of channels. This leads to "non-coherent" schemes that fail to exploit specific channel states, resulting in a performance floor.

This paper proposes a middle ground using meta-learning to find a "universal initialization" that allows for rapid adaptation to new, unseen channels.

System Model and Meta-Learning Approach

The authors utilize Model-Agnostic Meta-Learning (MAML) to optimize the initial weights of the communication system. Instead of minimizing error across all channels simultaneously, the meta-learning objective is to find a starting point $\theta$ such that a small number of gradient descent steps on a specific channel $h$ leads to optimal performance.

The end-to-end communication system modeled as an autoencoder with a transmitter, channel, and receiver.

The architecture, as shown in Figure 1, includes a feed-forward neural network for the transmitter, a normalization layer to enforce power constraints, and a receiver network with a softmax output. To enable backpropagation through the stochastic channel, the "reparameterization trick" is employed, allowing gradients to flow from the receiver back to the transmitter.

Experimental Results

The proposed method was evaluated on simulated Rayleigh block fading channels. The meta-training phase used 100 different channel realizations to learn the initialization, which was then tested on 20 entirely new channels.

Block error rate (BLER) vs. training iterations, comparing meta-learning, joint training, and fixed initialization.

As illustrated in Figure 3, the meta-learning approach demonstrates a dramatic speedup. While a standard "Fixed Initialization" requires over 1,000 iterations to reach a Block Error Rate (BLER) of $10^{-3}$, the meta-learned model achieves this in just 5 iterations. Furthermore, meta-learning successfully breaks the performance floor encountered by "Joint Training," which struggles to adapt to the specific nuances of individual channel states.

Conclusion

By applying MAML to the physical layer, this work shows that communication systems can be "pre-tuned" for rapid adaptation. This approach enables coherent-like performance with the speed of non-coherent schemes, requiring as little as one gradient step to adapt to new fading conditions. Future work will look toward model-free meta-learning to handle "black-box" channels where a differentiable model is unavailable.

Summary updated

Abstract

When a channel model is available, learning how to communicate on fading noisy channels can be formulated as the (unsupervised) training of an autoencoder consisting of the cascade of encoder, channel, and decoder. An important limitation of the approach is that training should be generally carried out from scratch for each new channel. To cope with this problem, prior works considered joint training over multiple channels with the aim of finding a single pair of encoder and decoder that works well on a class of channels. As a result, joint training ideally mimics the operation of non-coherent transmission schemes. In this paper, we propose to obviate the limitations of joint training via meta-learning: Rather than training a common model for all channels, meta-learning finds a common initialization vector that enables fast training on any channel. The approach is validated via numerical results, demonstrating significant training speed-ups, with effective encoders and decoders obtained with as little as one iteration of Stochastic Gradient Descent.
BibTeX
@inproceedings{park2020meta,
  author = {Sangwoo Park and Osvaldo Simeone and Joonhyuk Kang},
  title = {{Meta-Learning to Communicate: Fast End-to-End Training for Fading Channels}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=2lEQhQxjiE},
}
Figure