CoDecoder neural network pairs with decoders for up to 124x faster error correction
Researchers at the University of Colorado Boulder (Yicheng Guang, Yuliang Xiong, Xun Gao and Yueqi Chen) describe CoDecoder, a small neural network that works alongside an existing decoder, predicts likely logical outcomes and routes harder cases to the conventional decoder. The team reports decoding failures reduced by 6.4% to 31.4% versus correlated matching on surface codes of distance 7 to 17, and by 22% to 50% versus full-DEM Relay-BP on two bivariate-bicycle (BB) codes. Speedups of up to 124-fold over existing methods are reported, along with faster processing than re-implemented versions of the Cascade neural decoder on both code families; a joint decision step cut failures a further 14% to 18% on BB72 codes. The work is an arXiv preprint and has not yet been peer reviewed; the article notes that comparisons with other neural decoders such as Cascade are harder because training datasets differ.