Fast decoder for generic quantum codes via machine learning

Decoders for quantum error-correcting codes must use minimal resources to correct errors faster than they appear. We reduce this space-time requirement using neural networks and evolutionary algorithms to construct quantum decoders for generic stabilizer codes. In offline training, a map from syndromes to errors is optimized based on a cost function of the error locations. The trained network is then used to decode the syndrome bits in constant time for a given physical noise model. These techniques are tested on several block codes, small low-density parity check (LDPC) codes, and topological codes in both noiseless and noisy measurement regimes, and the performance and resource use are measured.