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