Decoding Small Surface Codes with Feedforward Neural Networks

• Quantum computing is being explored for applications including quantum-physics simulation, cryptanalysis, and unstructured search.Source: §I · p. 1
• Noise in quantum operations requires active error correction, whose co-processor must infer corrections from measurement results.A slow co-processor may degrade quantum-computer performance.Source: §I · p. 1
• The paper introduces a feedforward neural-network decoder after reviewing error correction, decoding speed, and existing fast-decoding methods.Source: §I · p. 1

• An encoded code with distance d prevents any operation on fewer than d qubits from transforming one encoded state into another.Source: §II · p. 2
• Pauli operators provide a basis for errors, and correcting them on a subset of qubits permits correction of arbitrary errors on that subset.Source: §II · p. 2
• Stabilizer codes encode states in a shared +1 eigenspace, and stabilizer measurements let a classical co-processor decode errors, though logical errors can remain.Source: §II · p. 2
• Fault-tolerant operations keep physical errors likely to remain correctable, while repeated fault-tolerant syndrome measurements address time-dependent errors.Source: §II · p. 2
• A fault-tolerance scheme's threshold marks where its code family can suppress errors exponentially, while a code's pseudothreshold compares encoded and bare-operation accuracy.Source: §II · p. 2
• Surface codes use stabilizers on adjacent qubits in a planar square tiling, and their syndrome changes can be decoded by minimizing error-chain lengths with Blossom.The measurement layout uses weight-four and weight-two stabilizers, with ancillas enabling indirect measurement.Source: §II · pp. 2–3, Figure 1 · p. 3, Figure 2 · p. 3

• Decoding is needed frequently to calculate the bits fed forward after logical-qubit measurements.Source: §III · p. 3
• In three consecutive measurement rounds, data-qubit errors produce horizontally separated syndrome changes, while measurement errors produce vertically separated changes.Source: Figure 3 · p. 4
• Stopping computation while decoding lets errors build up, while continuing syndrome measurements increases the future decoding backlog.The passage gives about 10 µs as an example error-build-up time in current superconducting circuits and 800 ns per syndrome-measurement round.Source: §III · p. 4

• Overlapping recovery processes windows of about d cycles incrementally, forwarding syndromes matched to the time boundary into the next window.Source: §IV · p. 4
• Parallelized Blossom uses message passing among local processors and scales to large distances, while a hardware-friendly distance-3 heuristic does not readily extend to higher distances.Source: §IV · pp. 4–5
• A stochastic neural decoder samples recovery chains from a learned error distribution, but repeated sampling takes an unknown amount of time.It produced performance similar to MWPM for simple error models.Source: §IV · p. 5
• The paper proposes feedforward neural networks as a simpler machine-learning decoder intended to provide high performance in bounded time.Source: §IV · p. 5

V. NEURAL NETWORK DECODER

• The error decomposition assigns a syndrome-matching pure error to a simple decoder, so the neural network only needs to identify the logical Pauli.The stabilizer component can be chosen arbitrarily without affecting decoder performance.Source: §V · p. 5
• The classification labels are the four possible logical Paulis: identity, X, Y, or Z.Classification training aims to label known inputs so that unknown inputs can also be labeled correctly.Source: §V · p. 5
• The feed-forward network uses trainable weights and biases, with average cross-entropy as its cost function.The network’s outputs can be rounded to binary values and interpreted as class labels.Source: §V · pp. 5–6, Figure 4 · p. 6
• Training samples are drawn at a physical error probability where Blossom’s logical error rate is about 25%, with at most 10^6 samples for larger codes.For small codes, the full set of possible syndromes can be sampled.Source: §V · p. 6
• Decoder performance is evaluated by comparing logical error rates against Blossom and a partial lookup table.The partial lookup table returns the identity correction for syndromes absent from its training set.Source: §V · p. 6

VI. RESULTS

• For QEC tests, inputs use syndrome bits, while fault-tolerance inputs include d measurement rounds and readout-derived stabilizer eigenvalues; outputs use two or four nodes depending on the scenario.The study tests channel-capacity and depolarizing QEC models, plus fault-tolerance models with noisy data, measurements, gates, and operations.Source: §VI · pp. 6–7, Table I · p. 7
• Training-set coverage varied by error model and code distance, including very small shares of frequent syndromes in several larger-distance and fault-tolerance tests.For example, coverage for distance-7 depolarizing QEC was 3 × 10^-7%; fault-tolerance coverage ranged from 30.09% to 0.01% across the listed models.Source: §VI · pp. 7–8
• As code distance increases, the neural-network decoder remains similar to Blossom and substantially outperforms PLUT across channel-capacity and depolarizing QEC models.These results show correction of syndromes absent from training; PLUT performance becomes suboptimal at larger distances with limited training coverage.Source: Figure 5 · p. 8, §VI · p. 8, Figure 6 · p. 9
• At distance 7, unseen correctable weight-≤3 errors leave channel-capacity decoding slightly below Blossom; depolarizing decoding is slightly better despite a few missed weight-3 errors.Blossom treats Y errors as separate X and Z errors; training Y errors as weight-1 errors improves the neural decoder’s relative performance.Source: §VI · p. 8
• In small-distance fault-tolerance tests, all decoders perform similarly; channel-capacity and depolarizing need few training syndromes to match Blossom, while circuit noise is slightly better with more.The circuit model required more syndromes than the other fault-tolerance models.Source: §VI · p. 9

VII. DISCUSSION AND CONCLUSION

• Parallel evaluation of independent corrections and neurons, with logarithmic-depth parity and summation, underpins the hardware timing estimate.The estimate assumes FPGA XOR and simple arithmetic operations can be completed in one clock cycle.Source: §VII · pp. 9–11
• Figure 7 compares the neural network decoder with MWPM and a partial look-up table for distance-3 codes across three error models.The panels cover channel-capacity and depolarizing models with measurement errors, and a circuit error model.Source: Figure 7 · p. 10
• ∼100 ns is the estimated runtime for the largest network’s 19 serial steps, assuming each operation takes one clock cycle.The estimate includes two multiplications, 15 addition steps, and two sigmoid evaluations.Source: §VII · p. 11
• Future work includes measuring hardware communication overhead, extending to finite-window syndrome feedforward, and testing larger-distance and alternate codes.The paper identifies communication overhead as a factor to determine in hardware implementation.Source: §VII · p. 11
• Feedforward neural networks provide fast and accurate decoding for small surface codes in both quantum error correction and fault-tolerant operations.The authors expect the relatively low hardware requirements and anticipated runtime to support near-term use.Source: §VII · p. 11