Convolutional neural network based decoders for surface codes

1. Problem Statement

Quantum computers are susceptible to decoherence caused by environmental interactions. Quantum Error Correction Codes (QECC), specifically Surface Codes, are the leading solution for protecting logical qubits. However, decoding the error syndromes of surface codes using classical algorithms (like Minimum Weight Perfect Matching, MWPM) becomes computationally expensive as the code distance ($d$) increases, potentially slowing down quantum computation.

The core challenge addressed is:

• Scalability: Classical decoders struggle with execution time as the number of physical qubits grows.
• Noise Adaptability: Decoders must handle complex noise models, including measurement errors and gate imperfections, not just simple bit-flip errors.
• Training Data: As code distance increases, the number of possible error syndromes grows exponentially, requiring massive training datasets that are difficult to generate and process.

2. Methodology

The authors propose and evaluate High-Level Decoders (HLD) based on Convolutional Neural Networks (CNNs). The decoding process is split into two stages:

1. Simple Decoder: A naive algorithm that corrects local errors by connecting them to the nearest boundary, restoring the ancillary state but potentially creating logical errors.
2. High-Level Decoder (HLD): A neural network that takes the error syndrome as input and classifies whether the simple decoder's correction resulted in a logical error.

Key Technical Approaches:

• Input Representation: The error syndrome is represented as a $2d-1 \times 2d-1$ matrix where data qubits are 0 and measurement qubits are $\{1, -1\}$.
• Noise Models: The study tests two models:
  • Depolarising Noise: Only data qubit errors ($X, Y, Z$).
  • Depolarising + Measurement Errors: Includes imperfect measurements and multiple correction cycles (circuit-level noise simulation).
• Architecture Optimization:
  • Comparison between Feed-Forward Neural Networks (FFNN) and CNNs.
  • Implementation of Dilated Convolutions to increase the receptive field without increasing the number of trainable parameters, crucial for higher code distances.
• Training Strategy:
  • Investigation of training on datasets with varying error probabilities ($p$).
  • Data Augmentation driven by Explainability: Using saliency maps to identify specific failure cases (e.g., long error chains) and generating synthetic training samples to address them.
• Explainability: Application of Occlusion and GradCAM techniques to generate saliency maps, visualizing which parts of the input syndrome influence the network's decision.

3. Key Contributions

1. CNN vs. FFNN Performance: Demonstrated that CNNs outperform FFNNs for surface code decoding, particularly as code distance increases, due to their ability to capture spatial correlations in error syndromes.
2. Dilated Convolutions: Successfully applied dilated convolution layers to scale decoders to higher code distances ($d=11$) while reducing the total number of trainable parameters.
3. Training Probability Optimization: Found that training on datasets with higher error probabilities generally yields better generalization, as these datasets contain more statistically relevant, complex error syndromes. However, this requires a "pre-training" strategy (starting from weights trained on lower probabilities) to prevent training instability at very high error rates.
4. Explainability-Driven Augmentation: Introduced a novel technique where saliency maps are used to diagnose specific decoder failures (e.g., missing long error chains). The authors then generated an enhanced training set specifically containing these failure cases, significantly boosting performance.
5. Robustness to Measurement Errors: Showed that CNN-based HLDs adapt better to noise models including measurement errors compared to standard MWPM benchmarks, which often degrade in performance under such conditions.

4. Results

• Accuracy vs. Code Distance:
  • For small distances ($d=7$), CNN-based HLDs outperformed MWPM.
  • For larger distances ($d=9, 11$), standard HLDs initially underperformed MWPM because the training set was too small to cover rare, high-weight error syndromes.
  • Dilated Convolutions improved performance for $d=11$ by allowing the network to see larger error chains without exploding parameter counts.
• Noise Model Adaptation: The CNN decoder maintained high accuracy when measurement errors were introduced, whereas MWPM accuracy dropped significantly unless specifically tuned for 3D matching graphs.
• Impact of Augmentation: By adding synthetic samples of length-5 error chains (identified via saliency maps) to the training set, the decoder's accuracy improved by ~1% at $p=0.1$ and up to 2% at higher error rates. Crucially, the augmented model successfully decoded error syndromes that the standard model and MWPM failed to handle correctly.
• Saliency Analysis: The maps revealed that the network focuses on error signals near the center of the code and the ends of error chains. Failures often occurred when the network failed to correlate distant parts of a long error chain, a flaw corrected by the augmented training.

5. Significance and Conclusion

This work establishes that Convolutional Neural Networks are a viable and promising alternative to classical decoders for Surface Codes, offering:

• Constant Execution Time: Critical for real-time quantum error correction.
• Adaptability: The ability to learn complex noise models without manual algorithmic re-tuning.
• Interpretability: The use of explainable AI (XAI) not only builds trust in the "black box" nature of neural networks but provides a direct feedback loop for algorithmic improvement.

The study concludes that while scaling to very high distances remains challenging due to the exponential growth of error syndromes, techniques like dilated convolutions, strategic training on high-error datasets, and explainability-driven data augmentation provide a clear pathway to building robust, high-performance decoders for future quantum hardware.