Autoencoder in 6G | AI/ML for Wireless Networks

The autoencoder is one of the most exciting and forward looking concepts for how AI/ML will fundamentally reshape the physical layer of 6G networks. We will explore how autoencoder works in 6G communication chain and how it leverages benefits of AI/ML.

What is an Autoencoder in General?

It is a special type of artificial neural network used for unsupervised learning. Its primary goal is to learn an efficient, compressed representation (an “encoding”) of a set of data. It consists of two main parts viz. encoder & decoder as used in communication chain.

1. The Encoder: This part of the network takes the original, high dimensional input data and compresses it down to a smaller, low dimensional representation. This compressed form is often called the “bottleneck” or “latent space.”

2. The Decoder: This part takes the compressed representation from the encoder and tries to reconstruct the original input data as accurately as possible.

In this AI model, the entire network is trained by comparing the final output to the original input and trying to minimize the difference i.e. “reconstruction error”.

How Autoencoder works in 6G wireless system

Here autoencoder architecture is mapped directly onto a wireless communication link, creating a radical new approach called “end-to-end learning.”

Let us understand how autoencoder works in 6G context. The figure mentions basic communication chain similar to 6G wireless system to illustrate autoencoder concept.

Image Courtesy : Rohde and Schwarz

1. Transmitter as encoder : Input bits are represented by “S”. Encoder neural network learns best possible ways and transform these bits to complex signal output as “X”. This transmitter block consists of various sub-blocks such as channel coding, modulation and waveform shaping as per physical layer specifications defined for 6G wireless system. The output “X” is transmitted over the air.
2. Wireless channel : The data “X” goes through various impairments such as noise, fading, interference before it is received as “Y” at the receiver input.
3. Receiver as decoder : The distorted, noisy signal “Y” is converted to final output “S’”. Neural network having end to end knowledge of chain utilizes best possible ways using algorithms to reconstruct original data by minimizing error (i.e. difference between “S’” and “S”). Receiver consists of demodulation, channel equalization and decoding as defined in Physical (PHY) layer specifications.

End to end learning approach is being utilized which is powered by autoencoder architecture as described below.

• Instead of optimizing each signal processing block in isolation, the AI learns the entire communication chain at once. It is trained on data that includes the real-world imperfections of the wireless channel.
• The AI model can be trained not just on channel distortions, but also on the known imperfections of its own hardware such as power amplifier non linearities (AM/AM, AM/PM), phase noise etc.
• The neural network is not constrained by conventional communication theory. It is free to explore the entire mathematical space of possible solutions to find the one that minimizes the error between the input and output.
• The AI model replaces a long and complex chain of individually designed and optimized blocks.