quantispect-structure-aware-3d-cnn-predecoder

name

quantispect-structure-aware-3d-cnn-predecoder

description

Structure-aware lightweight 3D CNN pre-decoder for scalable surface code quantum error correction. Use when designing or benchmarking neural QEC decoders for surface codes, implementing spatio-temporal syndrome processing, replacing dense 3D convolutions with factorized branches, or co-designing lightweight pre-decoders with classical matching decoders. Trigger words: QuantiSpect, surface code neural decoder, 3D CNN QEC, FastHyperBlock, quantum error correction pre-decoder, spatio-temporal syndrome decoder, lightweight neural decoder.

QuantiSpect: Structure-Aware Lightweight 3D CNN Pre-Decoder for Surface Codes

Core Idea

Replace expensive dense 3D convolutions in a surface-code neural pre-decoder with FastHyperBlocks that factor the spatio-temporal syndrome volume into three lightweight, structure-aware branches. The resulting QuantiSpect model preserves (or improves) decoding accuracy while dramatically cutting parameters and MACs, and it pairs with a classical minimum-weight perfect matching (MWPM) decoder to handle residual errors.

Problem

• Dense 3D CNNs for surface-code decoding scale poorly: parameter count and compute grow rapidly with receptive field.
• Prior art (e.g., the Accurate baseline) used stacked standard 3×3×3 convolutions to obtain large receptive fields, yielding ~1.80 M parameters.
• For real-time QEC and large code distances, a smaller, faster neural pre-decoder with the same accuracy is needed.

Solution (arXiv:2607.18204, 2026-07-20)

1. Input/Output Tensor

Tensor	Shape	Meaning
Input	(B, 4, T, D, D)	Batch × 4 channels (residual syndromes: X-space, Z-space, X-time, Z-time) × syndrome rounds T = d_m × spatial lattice D × D
Output	(B, 4, T, D, D)	Per-voxel correction maps for space-like X/Z and time-like X/Z errors

The data pipeline and residual-syndrome construction are inherited from the Accurate baseline; only the CNN architecture changes.

2. QuantiSpect Architecture

Three stages:

1. Stem — Conv3D(4 → C, 3×3×3) → GroupNorm → GELU
2. Main body — N identical FastHyperBlock residual blocks
3. Head — GroupNorm → Conv3D → GELU → Conv3D

Default hyperparameters:

Parameter	Symbol	Value
Input / output channels	—	4 / 4
Hidden dimension	C	96
Mid dimension	C_mid	144
Mixing groups	G	6
Number of blocks	N	5
Gate reduction ratio	r	4
Dropout probability	p_drop	0.02
Total parameters	—	~0.663 M
Receptive field	R	13

3. FastHyperBlock

Each block decomposes a full 3D convolution into three parallel, structure-aware branches, then fuses them with a channel gate and residual connection.

Pre-expansion:

• GroupNorm → Conv3D 1×1×1 (C → C_mid) → GELU
• Spatial/temporal operations run in the larger C_mid space; residual path stays at C.

Three parallel branches:

Branch	Kernel	Role	Parameters
Spatial	(1, 3, 3) depthwise	Local planar syndrome patterns	C_mid × 9
Temporal	(3, 1, 1) depthwise	Cross-round dynamics, measurement faults	C_mid × 3
Mixed	(3, 3, 3) grouped (G groups)	Joint spatio-temporal correlations	C_mid × 27 / G

Fusion:

• Sum branch outputs → Conv3D 1×1×1 projection
• Squeeze-and-Excitation (SE) channel gate: global pooling → FC(C/r) → FC(C) → sigmoid → channel-wise rescale
• GroupNorm
• Residual connection back to the C-dimensional input

Receptive field:

R = 1 + 2 + 2N   (kernel-3 stem + two voxels per FastHyperBlock)

With N = 5, R = 13. Each FastHyperBlock contributes about 128 k parameters.

Key Results

Accuracy vs Efficiency

• QuantiSpect (R=13) matches the Accurate baseline at the same receptive field.
• ~2.71× parameter reduction: 0.663 M vs 1.80 M.
• ~2.84× MAC reduction per voxel.

Logical Error Rate Improvement

When QuantiSpect is used as a pre-decoder and residual errors are passed to PyMatching:

• Up to ~1.85× lower logical error rate than uncorrelated PyMatching at code distance d = 13, physical error p = 0.5%.
• Up to ~3.11× faster PyMatching decode at d = 23 (smaller matching graph after pre-decoding).

Scaled Variant: QuantiSpect-21

• N = 10 blocks → receptive field R = 21.
• Only 1.18 M parameters.
• Raises the circuit-level threshold to ~0.80% under the evaluated noise model.

When to Use

• Building a lightweight neural pre-decoder for surface or toric codes.
• Need to scale receptive field without dense 3D convolution cost.
• Want to accelerate classical MWPM/union-find by reducing the residual graph size.
• Co-designing latency-constrained QEC decoders for large code distances.
• Porting syndrome decoding to edge/neuromorphic accelerators where parameter count matters.

Pitfalls

• FastHyperBlock assumes spatial isotropy of the syndrome lattice; may need adaptation for anisotropic or color-code geometries.
• The SE gate and grouped convolutions require channel counts divisible by G and r; choose C and C_mid accordingly.
• Performance is validated for surface-code residual syndromes; re-train and re-calibrate for other code families or noise models.
• A matching decoder is still required for the residual error graph; QuantiSpect is a pre-decoder, not a standalone decoder.

Implementation Checklist

• Implement residual-syndrome extraction from syndrome history (4 channels: X-space, Z-space, X-time, Z-time).
• Build FastHyperBlock: pre-expansion, three depthwise/grouped branches, SE gate, GroupNorm, residual.
• Stack N blocks with stem/head; verify receptive field R = 1 + 2 + 2N.
• Train with cross-entropy on error-correction labels from simulated circuit-level noise.
• Evaluate logical error rate and compare against MWPM baseline.
• Benchmark inference latency and parameter count; iterate on C, C_mid, G, and N.

Related Skills

• [[neural-decoder-quantum-error-correction]] — General neural decoder patterns for topological codes
• [[ml-quantum-error-correction]] — ML-assisted QEC including DiffQEC and RL control
• [[quantum-error-correction-methods]] — Broader QEC patterns and code families
• [[adaptive-syndrome-skipping-surface-gkp]] — Syndrome adaptive gain control for surface/GKP codes
• [[real-time-qec-system-stack]] — Real-time QEC system architecture and latency budgets

Key Paper

• "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, Gui-Lu Long
  arXiv:2607.18204 [quant-ph], 2026-07-20

Session Notes

• Created from 2026-07-22 daily research cron job.
• Methodology extracted from the arXiv HTML version of the paper.