- Authors:
-
Duque,
V.
,
Lewandowsky,
J.
,
Adrat,
M.
,
Hardenbicker,
T.
,
Jax,
P.
- Book Title:
- ICMCIS
- Organization:
- IST Panel Office
- Status:
- accepted for publication
- Date:
-
2024
- Language:
- English
Abstract
Autoencoders are particularly interesting deep learning models for communications, as they resemble the ar- chitecture of a classical transmission system. Therefore, the end- to-end learning of communication systems became very popular recently. Several works explored the idea to learn the transmitter, i.e., an encoder neural network and the corresponding receiver (decoder neural network) jointly in an end-to-end manner. In this paper, we propose a novel hybrid autoencoder system that amends a conventional transmission chain. The autoencoder inputs the modulated signal of a conventional transmitter and adds desired features to it, e.g., a possibly constant signal envelope and an improved resistance against transmission errors. This signal enhancement is achieved by the choice of proper loss functions during training and by adding suitable regularization layers to the encoder. The decoder part at the receiver aims to reobtain the classically modulated signal, such that the proposed signal enhancing autoencoder is transparent to the conventional transmission system. Like this, the original transmitter and receiver systems can still be used, while adding novel features to them via the autoencoder. Our presented results show that the proposed signal enhancing autoencoder can improve existing communication schemes in versatile ways with very low efforts, just by the choices of application specific loss functions and proper autoencoder architectures. The proposed concept is par- ticularly useful in the military context, where communication systems are deployed in large numbers and typically have long utilization periods, but might need improvements or novel features over their lifetime.
Download
BibTeX
