2023
Published
Transformer-QEC
Transformer-QEC: Quantum Error Correction Code Decoding with Transferable Transformers
Quantum decoding.
Built to transfer.
The example is described below.
The transformer reads check results and proposes local corrections. A conventional decoder handles the remaining syndrome. Change the code distance to see how the inputs grow around the shared model.
A view of the paper’s architecture. It does not run a trained decoder or show learned attention weights. Transfer to a new code size still needs further training.
Decode, then transfer.
Use a transformer decoder and transfer learning to decode quantum error-correction codes across code distances.
How the method works.
Quantum error correction produces a pattern of check results, called a syndrome. A decoder uses this pattern to infer errors. As a code grows, the pattern grows too, and decoding becomes harder.
The transformer reads the full pattern and proposes local corrections. A conventional decoder then handles the remaining syndrome. The two stages are combined to infer the final logical error. Transfer learning lets a trained model serve as the starting point for another code distance, a measure of code protection.
For example, the paper starts from a model trained at code distance 5. It then adjusts that model for other distances with 10 training passes, compared with 100 passes when training from scratch. The model reuses what it learned, but still receives further training.
This example describes the paper’s training setup. It does not guarantee the same training saving for every code or noise model. See Section 4.1 of the paper for the transfer experiment, and Section 3 for the two-stage decoder.
Sources &
citation.
Fast ML for Science Workshop at ICCAD2023
2023 · Published
Fast ML for Science @ ICCAD2023 · arXiv:2311.16082
View the citation
@inproceedings{transformer-qec-2023,
title = {{Transformer-QEC: Quantum Error Correction Code Decoding with Transferable Transformers}},
author = {Hanrui Wang and Pengyu Liu and Kevin Shao and Dantong Li and Jiaqi Gu and David Z. Pan and Yongshan Ding and Song Han},
year = {2023},
booktitle = {Fast ML for Science Workshop at ICCAD2023},
url = {https://arxiv.org/abs/2311.16082},
note = {arXiv:2311.16082. Workshop paper, not an ICCAD main-conference paper.}
}