Pith. sign in

REVIEW 2 major objections 7 minor 1 cited by

A factorized 3D convolutional pre-decoder, QuantiSpect, matches a dense baseline's surface-code decoding accuracy and ~0.77% circuit-level threshold with about a third of the parameters, and a deeper variant raises the threshold to ~0.80%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:40 UTC pith:WDRFUSPR

load-bearing objection Useful engineering result, but the parity claim rests on an unverified baseline reimplementation and the end-to-end speedup is never measured. the 2 major comments →

arxiv 2607.18204 v1 pith:WDRFUSPR submitted 2026-07-20 quant-ph

QuantiSpect: A Structure-Aware Lightweight 3D CNN Pre-Decoder for Scalable Surface Code Quantum Error Correction

Pan Gao , Xu-Sheng Xu , Ji-Ze Han , Jing-Wei Wen , Ling Qian , Xu-Dong Lv , Run-Qing Zhang , Xiao-Xiao Hu
show 1 more author
Gui-Lu Long
This is my paper
classification quant-ph PACS 03.67.Pp
keywords quantum error correctionsurface codeneural pre-decoder3D CNNfactorized convolutiondepthwise separable convolutionlogical error ratecircuit-level threshold
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

QuantiSpect is a lighter neural 'pre-decoder' for the surface code, the component that cleans most local errors so a slower global matching decoder only sees a sparse residual syndrome. This paper argues that current dense 3D convolutions are overkill because surface-code error patterns are partially separable in space and time. The proposed FastHyperBlock splits each dense 3x3x3 convolution into a depthwise spatial branch, a depthwise temporal branch, and a grouped spatio-temporal branch, followed by a channel gate. The paper reports that this factorized design matches the dense baseline's circuit-level threshold (p_th ≈ 0.77%) and logical-error-rate performance at moderate to large code distances while using ~2.7x fewer parameters and ~2.8x fewer per-voxel operations, and that adding blocks enlarges the receptive field cheaply enough to lift the threshold to ~0.80%. If correct, this makes real-time hybrid decoding substantially cheaper to deploy.

Core claim

The central claim is that a structure-aware, factorized 3D CNN can replace the dense 3D CNN pre-decoder in the established two-stage decoding pipeline without losing decoding power. QuantiSpect-13 reaches the same receptive field R=13 as the paper's Accurate baseline with 0.663M parameters and 0.633M per-voxel MACs versus 1.80M and 1.797M, yet reproduces the baseline's Wilson-weighted circuit-level threshold p_th=0.769% and its logical-error-rate curves for d≥9; relative to the matching decoder used alone it cuts the logical error rate up to 1.85x at d=13, p=0.5% and speeds up the decoder up to 3.11x at d=23. A deeper QuantiSpect-21 (1.18M parameters, R=21) raises the threshold to p_th≈0.80%

What carries the argument

FastHyperBlock, a residual block that replaces one dense 3x3x3 convolution with three parallel factorized branches: a depthwise spatial branch (kernel 1x3x3) for planar data-qubit error patterns, a depthwise temporal branch (kernel 3x1x1) for cross-round measurement correlations, and a grouped spatio-temporal branch (kernel 3x3x3) for coupled circuit-level faults, followed by a 1x1 fusion, squeeze-and-excitation channel gate, GroupNorm, and a residual connection. The receptive field follows R=1+2+2N, so each added block widens the field by 2 at roughly 128k parameters instead of the ~442k a dense layer costs, making receptive-field growth linear in parameter count rather than cubic.

Load-bearing premise

The load-bearing premise is that the dense baseline CNN was trained and evaluated under exactly the same protocol as QuantiSpect; the paper asserts this in Section IV but supplies no baseline training configuration beyond a citation, so a fairness gap in that comparison would invalidate the parity claim.

What would settle it

Reproduce the dense Accurate baseline under the exact QuantiSpect protocol (d=d_m=13, 100 epochs, ~33.6M samples/epoch, Lion with EMA, 50,000 shots/basis, Wilson-weighted FSS) and compare the resulting p_th and LER at d=13, p=0.5%; if the reproduction diverges from the reported 0.769% threshold or error curves beyond the 95% confidence intervals, the central parity claim is undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Real-time decoding pipelines can carry a neural pre-decoder at ~2.7x lower parameter and ~2.8x lower per-voxel compute cost without sacrificing the circuit-level threshold or large-distance accuracy.
  • Because the architecture is fully convolutional and distance-free, a single checkpoint trained at d=13 applies at inference to all tested distances d=5..23 without retraining.
  • A larger receptive field is a cheap path to better decoding: QuantiSpect-21 raises p_th from ~0.77% to ~0.80% while staying below the R=13 baseline's parameter count.
  • The hybrid pre-decoder + global matching speedup grows with code distance (up to 3.11x at d=23), so the benefit appears exactly where the global decoder's workload is heaviest.
  • The factorized model's small-distance dip (ρ≈1.01 at d=5 vs 1.46 for the dense baseline) shows the advantage is distance-dependent; at large d the ordering is LER_QS ≈ LER_Accurate < LER_Fast.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same space-time separability prior could be tested on other topological codes with anisotropic error structure, such as color codes or twisted toric codes; the paper does not claim this, but the architecture's assumptions are general enough to make it a natural experiment.
  • The d=5 underperformance suggests the factorization is a liability when the syndrome volume is too small for the branches to specialize; a hybrid that uses dense convolutions on the first few rounds or at small d might recover that loss.
  • The near-ideal effective-distance slopes (α≈0.456 and 0.481 vs ideal 0.5) hint that residual logical errors are dominated by long-range correlations; routing only those residuals to a global solver could approach ideal scaling more closely.
  • Since training fixes the physical error rate near p≈0.006, hardware deployment would need error-rate calibration; a cheap robustness test is to train on a mixture of p values and check whether threshold and LER stay stable across operating points.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. The paper proposes QuantiSpect, a factorized lightweight 3D CNN pre-decoder for the rotated surface code, built as a drop-in replacement for the dense CNN backbone of Ref. [1]. Each residual block replaces a dense 3×3×3 convolution with three parallel branches — depthwise spatial (1×3×3), depthwise temporal (3×1×1), and grouped spatio-temporal (3×3×3) — followed by squeeze-and-excitation gating. The base variant QuantiSpect-13 matches the R=13 receptive field of the Accurate baseline with 0.663 M parameters and 0.633 M MACs/voxel, and the paper claims it matches Accurate's circuit-level threshold (p_th = 0.769% under Wilson-weighted FSS) and its logical error rate for d ≥ 9, while reducing LER by up to 1.85× over uncorrelated PyMatching at d=13, p=0.5% and speeding up the PyMatching stage by up to 3.11× at d=23. A deeper variant QuantiSpect-21 (R=21, 1.18 M parameters) is reported to raise the threshold to ≈0.80%. Simulations span d=5–23 with 50,000 shots per basis and Wilson confidence intervals.

Significance. If the central parity claims hold, QuantiSpect is a useful contribution to real-time decoding research: it demonstrates that a structure-aware factorized CNN can match a dense pre-decoder at the same receptive field with substantially fewer parameters and MACs, and that the receptive field can be cheaply enlarged. The paper's strengths are the extensive unified benchmark (ten distances, six physical error rates, 50k shots per basis), the Wilson-interval error reporting, the 80-configuration FSS robustness scan, the consistent parameter/MAC bookkeeping, and the public release of the QuantiSpect model and code. The main risk is the reproducibility of the baseline comparison: the Fast/Accurate baselines are asserted to be retrained under identical conditions but no checkpoints, training logs, or baseline configurations are provided. I agree with the stress-test concern that this is the load-bearing weak point of the paper. A second, smaller weakness is that the QuantiSpect-21 comparison confounds receptive-field growth with changes in training distance and learning rate.

major comments (2)
  1. [Sec. III.G / Sec. IV] The central parity claim (Eq. 20, Tables VII–VIII) is that QuantiSpect-13 matches the Accurate baseline in threshold and LER. This claim depends on Fast and Accurate having been faithfully retrained from Ref. [1] under identical conditions. The manuscript asserts this ('all experiments use identical conditions', 'we retain the full data processing pipeline, training methodology, and residual syndrome construction from Ref. [1]') but provides no baseline checkpoints, no training logs, no commit hash, and no baseline-specific hyperparameter configuration. The Hugging Face release is described as covering QuantiSpect only. Since the LER differences between QuantiSpect-13 and Accurate at several distances (e.g. 3.90e-3 vs 4.46e-3 at d=13; 9.40e-4 vs 7.20e-4 at d=21) are comparable in magnitude to sampling uncertainty, a subtle implementation difference in the baseline — optimizer schedule, H
  2. [Sec. V] The claim that the larger receptive field of QuantiSpect-21 'captures longer-range error correlations that lift the threshold' is confounded. The R=21 variant changes three things simultaneously relative to QuantiSpect-13: the number of blocks (N=5→9), the training distance and number of rounds (d=d_m=13→21), and the initial learning rate (2×10^-4→1×10^-4). Therefore the threshold improvement and LER reduction in Table XII cannot be attributed to R alone. To support the causal statement, the paper needs a control, e.g. QuantiSpect-21 trained at d=13, or QuantiSpect-13 trained at d=21, or a dense R=21 baseline trained under the same protocol. Without such a control, the receptive-field scaling conclusion is not established.
minor comments (7)
  1. [Table VIII / Eq. (20)] The 'matches Accurate' claim would be easier to assess if the table included confidence intervals or a statement of which differences are within Wilson intervals. At d=21 the QS-13 value (9.40e-4) is about 30% higher than Accurate (7.20e-4), while at d=23 it is slightly lower; a quantitative statement about statistical compatibility is needed.
  2. [Table XII / Sec. V.b] At d=5, QuantiSpect-21 has ρ=0.92, meaning the hybrid decoder is worse than PyMatching alone. The text says the two variants are 'similar at small distances'; this degradation should be noted explicitly since it is a real limitation of the pre-decoder at very small distances.
  3. [Table XI / Sec. IV.D] The conclusion that QuantiSpect has the fastest effective-distance growth is not statistically supported: α=0.456±0.016 for QS-13 overlaps with α=0.432±0.009 for Accurate, and α=0.481±0.012 for QS-21 also overlaps with the baseline values. Please soften 'the steepest slope' and 'sits closest to ideal' in the conclusion, or add a test of significance.
  4. [Sec. IV.A / Appendix B] The quoted thresholds p_th=0.769±0.002 carry only curve-fit covariance. The robustness scan shows systematic ranges of 0.60–0.78% (QuantiSpect) and 0.67–0.78% (Accurate); this systematic sensitivity is much larger than ±0.002 and should be reported in the main text next to Table VII so readers do not over-interpret the precision.
  5. [Sec. IV] The basis-averaged Wilson interval is computed as the average of the separate lower and upper bounds for X and Z bases. This is not an exact confidence interval for the average of two binomial proportions. Please either use a combined binomial interval or state explicitly that the averaged interval is approximate.
  6. [Sec. III.A / III.E] The architecture description does not specify padding and stride for the stem and block convolutions. Since the stated tensor shape (B,C,T,D,D) is preserved throughout, padding must be used; please state the padding convention explicitly.
  7. [Miscellaneous] There are several typos and formatting issues: 'standaalone' and 'codition' in Sec. VI, 'F ast' spacing throughout, 'V ertical' in Fig. 4, and 'Notice that' in Table VI. A final copyedit is needed.

Circularity Check

0 steps flagged

No significant circularity: all headline claims are measured against an external baseline and no parameter is fitted to a target quantity.

full rationale

The paper's central claims are empirical benchmark results, not derivations from fitted constants. Thresholds and logical error rates come from Stim circuit-level simulations (Sec. IV, N=50000 shots per basis) and are processed through a finite-size-scaling ansatz (Eq. 19); the fitted p_th is a measured property of the simulated decoder, not an input that the architecture was tuned to reproduce. The comparison with Fast/Accurate uses the external Ref. [1] pipeline (data generation, homological-equivalence labels, residual construction), and Ref. [1] has no author overlap with the current paper, so this is not a self-citation chain. The architecture-level numbers (R=1+2+2N, parameter/MAC counts in Tables III-V) follow by arithmetic from the stated kernel sizes and channel counts. The claim of matching Accurate is a controlled benchmark that depends on an asserted identical pipeline; whether the baselines were faithfully reproduced is a reproducibility concern (no checkpoints/training configs are provided), not circularity. The paper itself flags statistical limitations (Sec. IV A systematic fitting-window sensitivity; Sec. V low failure counts at large d; Appendix B exploratory event-count check), which show care but do not make any claim true by definition. No equation in the paper reduces a predicted quantity to the quantity it is supposed to predict.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central empirical claim rests on standard QEC background, the Ref [1] label/residual pipeline, and hand-chosen architecture hyperparameters. No new physical entity or fitted constant is introduced. The main unverified imports are the homological-equivalence label protocols and the exact baseline implementation; the FSS ansatz and fitting window add fitting freedom.

free parameters (3)
  • Architecture hyperparameters (C=96, Cmid=144, G=6, r=4, N=5/9, dropout=0.02) = C=96, Cmid=144, G=6, r=4, N=5/9, p_drop=0.02
    Hand-chosen to match the R=13 baseline and control parameter budget; no systematic ablation is reported, so the efficiency ratios may shift with different choices.
  • Training noise upscale p_train = ~0.006
    Training data are generated at one effective physical error rate near threshold, while evaluation sweeps p=0.1% to 0.8%. Transfer across rates is assumed, not fitted.
  • Logit binarization threshold = 0.5
    A fixed sigmoid threshold is used to convert network outputs into binary corrections; it is not tuned, and results could change with a calibrated threshold.
axioms (5)
  • domain assumption Circuit-level depolarizing noise model of Ref [1] is a valid proxy for hardware noise
    Used to generate all training and evaluation data; if the noise model is unrepresentative, the reported thresholds and logical error rates do not transfer to real hardware.
  • standard math Stim DEM construction and PyMatching produce a correct MWPM for independent faults
    The whole pipeline composes detection events into a weighted matching graph using the standard DEM formalism; this is unverified background but standard in the field.
  • domain assumption Spacelike/timelike homological-equivalence label protocols from Ref [1] produce unambiguous, correct labels
    The network training labels are canonicalized by these external algorithms; the paper does not re-implement or verify them, and they are load-bearing for training quality.
  • domain assumption Finite-size scaling ansatz Eq. (19) with polynomial order 2 describes the threshold
    Used to extract pth for all methods over p in [0.3%,0.8%] and d>=7, excluding zero-LER points; the robustness scan shows sensitivity to this choice.
  • domain assumption A fully convolutional network trained at d=13 generalizes to all distances and round counts at inference
    Assumed because there are no distance-dependent parameters; the reported results at large d support but do not prove this assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 28514 in / 13590 out tokens · 117680 ms · 2026-08-01T15:40:46.997381+00:00 · methodology

0 comments
read the original abstract

Real-time decoding is a critical bottleneck for large-scale fault-tolerant quantum computing. AI-based neural pre-decoders locally correct most physical errors before passing residual syndromes to a global decoder, enabling sub-microsecond latencies. However, existing architectures carry significant overhead from dense 3D convolutions. We present QuantiSpect, a lightweight 3D convolutional neural network (CNN) pre-decoder for the rotated surface code, built on the decoding pipeline of Chamberland et al. The key idea is to replace the dense 3D convolutions with three parallel branches in each residual block: a depthwise spatial branch, a depthwise temporal branch, and a grouped spatio-temporal branch, followed by a squeeze-and-excitation channel gate. This reflects the structure of surface code errors, where spatial and temporal syndrome correlations are partially separable. On a unified 4xA100 GPU benchmark, QuantiSpect matches the receptive field of the Accurate baseline at R=13 while using ~2.71x fewer parameters (0.663M vs 1.80M) and ~2.84x fewer per-voxel convolutional MACs. It matches Accurate's circuit-level threshold and accuracy at moderate and large code distances, reduces the logical error rate by up to ~1.85x relative to uncorrelated PyMatching at d=13, p=0.5%, and speeds up the PyMatching decode by up to 3.11x at d=23. We also explored enlarging the receptive field by adding blocks. Even at R=21, the model uses only 1.18M parameters, fewer than both the R=13 Accurate baseline (1.80M) and the R=17 dense model (4.22M), despite its larger receptive field. This expanded variant significantly outperforms the Accurate model, raising the circuit-level threshold to ~0.80% and further reducing the logical error rate. Together, both variants show that a structure-aware factorized design is an effective, parameter-efficient alternative to a dense one for decoding the surface code.

Figures

Figures reproduced from arXiv: 2607.18204 by Gui-Lu Long, Jing-Wei Wen, Ji-Ze Han, Ling Qian, Pan Gao, Run-Qing Zhang, Xiao-Xiao Hu, Xu-Dong Lv, Xu-Sheng Xu.

Figure 1
Figure 1. Figure 1: FIG. 1. Two-stage decoding workflow [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: shows a d = 7 example. Data qubits (black circles) sit on the vertices of a d×d square lattice. Ancilla qubits (open circles) sit at the centers of the plaquettes and are used to measure the stabilizer operators. The X-type and Z-type stabilizers are arranged in a checker￾board pattern. Each X-type stabilizer (orange) is a prod￾uct of Pauli-X operators on the data qubits at the corners of its plaquette, ea… view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Syndrome extraction circuits for bulk (weight-four) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. The space–time syndrome volume for the rotated [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. End-to-end hybrid decoding pipeline. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Architecture of the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Logical error rate vs. physical error rate for each [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Head-to-head comparison of logical error rate at six representative code distances, panels (a)–(f) for [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Head-to-head logical error rate comparison across all [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Finite-size scaling data collapse for the three pre-decoder methods. Panels (a)–(c) use the unweighted fit; panels [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. QAdapt: A Noise-Adaptive Neural Pre-Decoding Framework for Quantum Error Correction

    cs.LG 2026-07 conditional novelty 5.0

    A continually adapted neural pre-decoder reduces logical error rate and residual matching latency versus a fixed neural baseline across 110 OOD noise settings and zero-shot on Willow.

Reference graph

Works this paper leans on

45 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Stem The stem maps the four-channel input to aC- dimensional hidden space: Stem : Conv3D(4→C, k= 3)→GN(C)→GELU, (9) 9 QuantiSpect Neural Pre-Decoder Architecture Main Pipeline Input Syndrome Volume (B,4,T,D,D) Stem Conv3D 4→96, kernel 3×3×3 GroupNorm(8 groups), GELU (B,96,T,D,D) Main Body FastHyperBlock ×5 detail Head GroupNorm(8 groups) Conv3D 96→96, ker...

  2. [2]

    Each block preserves the tensor shape (B, C, T, D, D) and is the core architec- tural contribution of this work

    FastHyperBlock The main body consists ofN= 5 identical FastHyperBlockresidual blocks. Each block preserves the tensor shape (B, C, T, D, D) and is the core architec- tural contribution of this work. a. Pre-expansion.The input is expanded fromCto Cmid channels via GN→Conv3D 1×1×1(C→C mid)→ GELU. Following the inverted bottleneck principle [25], spatial and...

  3. [3]

    The final convolution has no activation, since the outputs are raw logits for the four correction channels (spacelikeZ/Xand timelike X/Z)

    Head The head maps theC-dimensional features back to four output channels via GN→Conv3D 1×1×1(C→C)→ GELU→Conv3D 1×1×1(C→4). The final convolution has no activation, since the outputs are raw logits for the four correction channels (spacelikeZ/Xand timelike X/Z). D. Design rationale and comparison The three-branch design ofQuantiSpectis moti- vated by the ...

  4. [4]

    per-round

    yield ˆcZ,ˆcX ∈ {0,1}T×D×D , representing estimated data-qubit Pauli corrections at each round. The timelike channels (channels 3–4) yield ˆm X ,ˆmZ ∈ {0,1}T×D×D , representing estimated measurement bit-flip corrections. b. Induced syndrome via parity check.The space- like corrections are mapped to the stabilizer basis using the parity-check matrices. For...

  5. [5]

    Wilson-weighted fits Table XIV repeats the FSS fit for every method un- der both weighting schemes, includingQuantiSpect-21 (Sec

    Unweighted vs. Wilson-weighted fits Table XIV repeats the FSS fit for every method un- der both weighting schemes, includingQuantiSpect-21 (Sec. V). Under the unweighted fit, every method yields a lower threshold than under Wilson weighting, by 0.03–0.06 per- centage points, but the collapse RMSE is roughly 2× tighter (0.14–0.17 vs. 0.30–0.43 in log 10 sp...

  6. [6]

    Unwt” uses raw LER values; “Wilson

    Event-count-based robustness check As a further check, we also refit each method with an unweighted least-squares fit after excluding points with fewer than a minimum number of estimated observed logical-failure events, following the event-count filtering approach used in some surface-code threshold studies [38] as an alternative to explicit statistical w...

    2000

  7. [7]

    IV A are generated as the Cartesian product of the following parameter axes: •Polynomial orderoff:{2,3}

    Robustness scan configurations The 80 fitting configurations used in the robustness scan of Sec. IV A are generated as the Cartesian product of the following parameter axes: •Polynomial orderoff:{2,3}. •Error-rate window[p min, pmax]: full range, (0.2%,0.7%), (0.3%,0.7%), (0.3%,0.8%), (0.2%,0.8%). •Minimum distanced min: all (d≥5), 7, 9, 11. •W eighting: ...

    2025

  8. [8]

    Fast and accurate ai- based pre-decoders for surface codes.arXiv preprint arXiv:2604.12841, 2026

    Christopher Chamberland, Jan Olle, Muyuan Li, Scott Thornton, and Igor Baratta. Fast and accurate ai- based pre-decoders for surface codes.arXiv preprint arXiv:2604.12841, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Scheme for reducing decoherence in quan- tum computer memory.Physical review A, 52(4):R2493, 1995

    Peter W Shor. Scheme for reducing decoherence in quan- tum computer memory.Physical review A, 52(4):R2493, 1995

    1995

  10. [10]

    Error correcting codes in quantum theory.Physical Review Letters, 77(5):793, 1996

    Andrew M Steane. Error correcting codes in quantum theory.Physical Review Letters, 77(5):793, 1996

    1996

  11. [11]

    California Institute of Technology, 1997

    Daniel Gottesman.Stabilizer codes and quantum error correction. California Institute of Technology, 1997

    1997

  12. [12]

    Resilient quantum computation.Science, 279(5349):342–345, 1998

    Emanuel Knill, Raymond Laflamme, and Wojciech H Zurek. Resilient quantum computation.Science, 279(5349):342–345, 1998

    1998

  13. [13]

    Quantum error correction for quan- tum memories.Reviews of Modern Physics, 87(2):307– 346, 2015

    Barbara M Terhal. Quantum error correction for quan- tum memories.Reviews of Modern Physics, 87(2):307– 346, 2015

    2015

  14. [14]

    Fault-tolerant quantum computation by anyons.Annals of physics, 303(1):2–30, 2003

    A Yu Kitaev. Fault-tolerant quantum computation by anyons.Annals of physics, 303(1):2–30, 2003

    2003

  15. [15]

    Quantum codes on a lattice with boundary.arXiv preprint quant-ph/9811052, 1998

    Sergey B Bravyi and A Yu Kitaev. Quantum codes on a lattice with boundary.arXiv preprint quant-ph/9811052, 1998

    Pith/arXiv arXiv 1998

  16. [16]

    Topological quantum memory.Journal of Math- ematical Physics, 43(9):4452–4505, 2002

    Eric Dennis, Alexei Kitaev, Andrew Landahl, and John Preskill. Topological quantum memory.Journal of Math- ematical Physics, 43(9):4452–4505, 2002

    2002

  17. [17]

    Surface code quantum computing by lattice surgery.New Journal of Physics, 14(12):123011, 2012

    Dominic Horsman, Austin G Fowler, Simon Devitt, and Rodney Van Meter. Surface code quantum computing by lattice surgery.New Journal of Physics, 14(12):123011, 2012

    2012

  18. [18]

    A game of surface codes: Large-scale quantum computing with lattice surgery.Quantum, 3:128, 2019

    Daniel Litinski. A game of surface codes: Large-scale quantum computing with lattice surgery.Quantum, 3:128, 2019

    2019

  19. [19]

    Low overhead quan- tum computation using lattice surgery.arXiv preprint arXiv:1808.06709, 2018

    Austin G Fowler and Craig Gidney. Low overhead quan- tum computation using lattice surgery.arXiv preprint arXiv:1808.06709, 2018

    Pith/arXiv arXiv 2018

  20. [20]

    Real- time decoding for fault-tolerant quantum computing: Progress, challenges and outlook.Nano Futures, 7(3):032003, 2023

    Francesco Battistel, Christopher Chamberland, Kauser Johar, Ramon WJ Overwater, Fabio Sebastiano, Luka Skoric, Yosuke Ueno, and Muhammad Usman. Real- time decoding for fault-tolerant quantum computing: Progress, challenges and outlook.Nano Futures, 7(3):032003, 2023

    2023

  21. [21]

    Sparse blossom: cor- recting a million errors per core second with minimum- weight matching.Quantum, 9:1600, 2025

    Oscar Higgott and Craig Gidney. Sparse blossom: cor- recting a million errors per core second with minimum- weight matching.Quantum, 9:1600, 2025

    2025

  22. [22]

    Almost-linear time decoding algorithm for topological codes.Quantum, 5:595, 2021

    Nicolas Delfosse and Naomi H Nickerson. Almost-linear time decoding algorithm for topological codes.Quantum, 5:595, 2021

    2021

  23. [23]

    Neural decoder for topological codes.Physical review letters, 119(3):030501, 2017

    Giacomo Torlai and Roger G Melko. Neural decoder for topological codes.Physical review letters, 119(3):030501, 2017

    2017

  24. [24]

    Learning high-accuracy error decod- ing for quantum processors.Nature, 635(8040):834–840, 2024

    Johannes Bausch, Andrew W Senior, Francisco JH Heras, Thomas Edlich, Alex Davies, Michael Newman, Cody Jones, Kevin Satzinger, Murphy Yuezhen Niu, Sam Blackwell, et al. Learning high-accuracy error decod- ing for quantum processors.Nature, 635(8040):834–840, 2024

    2024

  25. [25]

    Scal- able neural decoder for topological surface codes.Physi- cal Review Letters, 128(8):080505, 2022

    Kai Meinerz, Chae-Yeun Park, and Simon Trebst. Scal- able neural decoder for topological surface codes.Physi- cal Review Letters, 128(8):080505, 2022

    2022

  26. [26]

    A scalable and fast artificial neural network syn- drome decoder for surface codes.Quantum, 7:1058, 2023

    Spiro Gicev, Lloyd CL Hollenberg, and Muhammad Us- man. A scalable and fast artificial neural network syn- drome decoder for surface codes.Quantum, 7:1058, 2023

    2023

  27. [27]

    Transformer-qec: Quantum error correction code de- coding with transferable transformers.arXiv preprint arXiv:2311.16082, 2023

    Hanrui Wang, Pengyu Liu, Kevin Shao, Dantong Li, Ji- aqi Gu, David Z Pan, Yongshan Ding, and Song Han. Transformer-qec: Quantum error correction code de- coding with transferable transformers.arXiv preprint arXiv:2311.16082, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Decoding surface code with a distributed neural network–based decoder.Quantum Machine Intelligence, 2(1):3, 2020

    Savvas Varsamopoulos, Koen Bertels, and Carmen G Al- mudever. Decoding surface code with a distributed neural network–based decoder.Quantum Machine Intelligence, 2(1):3, 2020

    2020

  29. [29]

    Neural network decoders for large-distance 2d toric codes.Quantum, 4:310, 2020

    Xiaotong Ni. Neural network decoders for large-distance 2d toric codes.Quantum, 4:310, 2020

    2020

  30. [30]

    Identity mappings in deep residual networks

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630–645. Springer, 2016

    2016

  31. [31]

    Mobilenets: Efficient con- volutional neural networks for mobile vision applications

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. Mobilenets: Efficient con- volutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017

    Pith/arXiv arXiv 2017

  32. [32]

    Mobilenetv2: In- verted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: In- verted residuals and linear bottlenecks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4510–4520, 2018

    2018

  33. [33]

    Opti- mal resources for topological two-dimensional stabilizer codes: Comparative study.Physical Review A—Atomic, Molecular, and Optical Physics, 76(1):012305, 2007

    H´ ector Bomb ´ ın and Miguel A Martin-Delgado. Opti- mal resources for topological two-dimensional stabilizer codes: Comparative study.Physical Review A—Atomic, Molecular, and Optical Physics, 76(1):012305, 2007

    2007

  34. [34]

    Low-distance surface codes under realistic quantum noise.Physical Review A, 90(6):062320, 2014

    Yu Tomita and Krysta M Svore. Low-distance surface codes under realistic quantum noise.Physical Review A, 90(6):062320, 2014

    2014

  35. [35]

    Stim: a fast stabilizer circuit simulator

    Craig Gidney. Stim: a fast stabilizer circuit simulator. Quantum, 5:497, 2021

    2021

  36. [36]

    Surface codes: To- wards practical large-scale quantum computation.Phys- ical Review A—Atomic, Molecular, and Optical Physics, 86(3):032324, 2012

    Austin G Fowler, Matteo Mariantoni, John M Mar- tinis, and Andrew N Cleland. Surface codes: To- wards practical large-scale quantum computation.Phys- ical Review A—Atomic, Molecular, and Optical Physics, 86(3):032324, 2012

    2012

  37. [37]

    Confinement-higgs transition in a disordered gauge the- ory and the accuracy threshold for quantum memory.An- 22 nals of Physics, 303(1):31–58, 2003

    Chenyang Wang, Jim Harrington, and John Preskill. Confinement-higgs transition in a disordered gauge the- ory and the accuracy threshold for quantum memory.An- 22 nals of Physics, 303(1):31–58, 2003

    2003

  38. [38]

    Suppressing quantum errors by scaling a surface code logical qubit.Nature, 614(7949):676–681, 2023

    Google Quantum AI. Suppressing quantum errors by scaling a surface code logical qubit.Nature, 614(7949):676–681, 2023

    2023

  39. [39]

    Surface code quantum computing with error rates over 1%.Physical Review A—Atomic, Molecular, and Optical Physics, 83(2):020302, 2011

    David S Wang, Austin G Fowler, and Lloyd CL Hollen- berg. Surface code quantum computing with error rates over 1%.Physical Review A—Atomic, Molecular, and Optical Physics, 83(2):020302, 2011

    2011

  40. [40]

    Designing fault-tolerant cir- cuits using detector error models.Quantum, 9:1905, 2025

    Peter-Jan HS Derks, Alex Townsend-Teague, Ansgar G Burchards, and Jens Eisert. Designing fault-tolerant cir- cuits using detector error models.Quantum, 9:1905, 2025

    1905

  41. [41]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. InProceedings of the IEEE conference on com- puter vision and pattern recognition, pages 7132–7141, 2018

    2018

  42. [42]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vi- sion (ECCV), pages 3–19, 2018

    2018

  43. [43]

    Symbolic discovery of optimization algorithms.Advances in neural information processing systems, 36:49205–49233, 2023

    Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, et al. Symbolic discovery of optimization algorithms.Advances in neural information processing systems, 36:49205–49233, 2023

    2023

  44. [44]

    Probable inference, the law of succes- sion, and statistical inference.Journal of the American Statistical Association, 22(158):209–212, 1927

    Edwin B Wilson. Probable inference, the law of succes- sion, and statistical inference.Journal of the American Statistical Association, 22(158):209–212, 1927

    1927

  45. [45]

    Fault-tolerant thresholds for quan- tum error correction with the surface code.Physical Re- view A, 89(2):022321, 2014

    Ashley M Stephens. Fault-tolerant thresholds for quan- tum error correction with the surface code.Physical Re- view A, 89(2):022321, 2014

    2014