Optimal soft-information inputs for the neural-network decoder

Determine whether providing defect probabilities derived from IQ readout and leakage flags as inputs is the optimal way to represent soft readout information to the recurrent neural network decoder used for the Surface-13 distance-three bit-flip surface-code experiment, with the goal of maximizing logical fidelity and minimizing the extracted logical error rate.

Background

The paper demonstrates that incorporating analog (soft) readout information into decoding improves logical performance in a superconducting implementation of a distance-three bit-flip surface code (Surface-13). Two decoders are studied: minimum-weight perfect matching (MWPM) and a recurrent neural network (NN).

For the NN, the authors supply soft information via defect probabilities computed from IQ readout values and leakage flags indicating occupation of the transmon’s |2⟩ state. While this approach yields improved performance over using hardened (binary) defects alone, the authors explicitly note uncertainty about whether this is the best way to present soft information to the network, motivating an investigation into optimal input encodings for NN decoders.