.png){kind=icon}
Return
U-Shaped Error Correction Code Transformers
.png){kind=icon}
PRE
AI
DOI：10.1109/TCCN.2024.3482349.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this work, we introduce two variants of the U-shaped error correction code transformer (U-ECCT) in combination with weight-sharing to improve the decoding performance of the error correction code transformer (ECCT) for moderate-length linear codes. The proposed models are inspired by the well-known U-Net architecture to leverage residual information for faster error estimation based on the syndrome-based reliability decoding principle. As an effort to further improve the general decoding performance of the U-ECCT, we propose the variational U-ECCT (VU-ECCT), in which the process of learning the shortcut connections is treated as a generative problem, forming a variational autoencoder (VAE) that exists intertwined with the existing U-ECCT model. This design allows the extraction of mutual information between the different levels of the U-shaped architecture, thus enhancing the performance of large syndrome sequences for low-rate codes. Additionally, to further reduce the model size, a new weight-sharing strategy, called mirror-sharing, is proposed to compress the model size as well as complement the mechanism of the proposed U-shaped architecture. In experiments, it has been demonstrated that our proposed models achieve significantly better performance than baseline conventional algorithms and other learning-based models.
Keywords:
Transformer
error correction code transformer (ECCT)
mirror-sharing
variational autoencoder (VAE)
U-Net
Journal
I
IF:
7
Papers:
1.6K
Citations:
5.5K
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSaveReduced complexity iterative decoding of low-density parity check codes based on belief propagation
PREAI
ShareSaveShareSave
View more