Neural-Network Decoders
Machine-learning decoders learn to map syndromes to corrections from data, capturing correlated noise that hand-built decoders assume away.
Decoding as a learned map
A decoder's job is to turn a syndrome into the most likely correction. Classical algorithms like matching assume a simple, independent noise model. A neural-network decoder instead learns the syndrome-to-correction map from examples, either simulated or measured, so it can absorb the real device's correlated and non-uniform noise without an explicit model.
Architectures
Because the surface-code lattice is a grid, convolutional networks are a natural fit, and recurrent or transformer models handle the time dimension of repeated syndrome rounds. Some designs predict the logical correction directly; others predict a residual on top of a fast classical decoder, letting the network fix only the cases where the classical decoder is likely wrong.
- Learns from data, so it can model leakage, crosstalk, and drift.
- Grid structure suits convolutional and graph neural networks.
- Recurrent or transformer layers process multi-round syndrome histories.
- Training requires representative data and must be refreshed as hardware drifts.
Demonstrations have shown learned decoders matching or beating matching-based decoders on real superconducting-device data, precisely because they exploit correlations that the idealized noise model ignores. The gains are largest where the true noise is far from the independent-error assumption.
The open issues are generalization and speed. A network trained on one code distance or one device may not transfer to another, and inference must finish inside the tight real-time latency budget. Recent work compiles trained decoders to specialized hardware to meet the microsecond-scale deadlines that fast qubits impose.