Neural network decoder confidence as a learned proxy for the logical gap

This paper demonstrates that the logit output of a graph neural network decoder trained on syndrome data serves as a superior learned proxy for the logical gap compared to the traditional minimum-weight perfect matching confidence measure, enabling more effective confidence-based post-selection and lower logical error rates in quantum error correction.

The Gist

Imagine you are trying to solve a massive, complex puzzle (a quantum computer's error correction) while wearing blindfolded gloves. You can't see the whole picture, only small clues (called "syndromes") that pop up on a screen. Your job is to guess which piece fits where to fix the puzzle.

Sometimes, you are right; sometimes, you are wrong. The big question is: How can you tell if your guess is a lucky guess or a solid, reliable one?

This paper is about teaching a computer to not just make a guess, but to say, "I'm 90% sure this is right," or "I'm only 50% sure." The authors wanted to see if a smart computer program (a Neural Network) could learn to give these "confidence scores" better than the traditional math tools used by scientists.

Here is the breakdown of their findings using simple analogies:

1. The Two Competitors: The "Math Rulebook" vs. The "Smart Student"

• The Math Rulebook (MWPM): This is the old-school method. It works like a strict accountant. It calculates the "distance" between errors and picks the shortest path to fix them. It has a built-in way to measure confidence called the "Logical Gap." Think of this as a ruler: if the gap between the best path and the second-best path is huge, the accountant is very confident. If the gap is tiny, it's unsure.
• The Smart Student (GNN): This is a Neural Network. It doesn't use a ruler or a rulebook. Instead, it was trained by looking at millions of examples of puzzles and their solutions. It learned to recognize patterns intuitively, like a student who has studied hard for a test. When it makes a guess, it outputs a "logit" (a number) that acts as its confidence score.

2. The Big Test: Who is Better at Filtering Mistakes?

The researchers wanted to see which method was better at Post-Selection. Imagine you are a teacher grading a test. You can throw away the answers you are least sure about to make sure your final grade is perfect.

• The Goal: Throw away the "maybe" answers and keep only the "definitely right" ones.
• The Result: The "Smart Student" (GNN) was much better at this. When they used the GNN's confidence score to decide which answers to keep, the final error rate was lower than when they used the Math Rulebook's ruler.

The Analogy:
Imagine the Math Rulebook is a security guard who stops people based on a strict height requirement. It's good, but it misses some bad guys who are just slightly shorter than the limit.
The Smart Student is a security guard who looks at your whole face, your walk, and your vibe. It turns out, the Student is better at spotting the "imposter" answers and keeping the "honest" ones, even if the Student can't explain exactly why using a ruler.

3. What Did They Find?

• The "Gap" is Real: Even though the Smart Student wasn't taught how to use a ruler, it naturally learned to act like one. When the Student was very confident, it was usually right. When it was unsure, it was usually wrong.
• The "Super-Confident" Tail: The Student had a special trick. For the answers it got right, it gave them huge confidence scores (like shouting, "I'm 100% sure!"). The Math Rulebook was more conservative; it rarely gave scores that high, even when it was right. This allowed the researchers to keep more of the "good" answers while still throwing away the "bad" ones.
• Calibration: The researchers checked if the confidence numbers actually matched reality. If the Student said "90% chance of being right," was it actually right 90% of the time?
  • The Math Rulebook was a bit off (it was slightly overconfident or underconfident depending on the situation).
  • The Smart Student was much closer to the truth. Its confidence numbers were a more accurate reflection of reality.

4. Why Does This Matter?

The paper concludes that you don't need to be a mathematician to get a good confidence score. You can just train a neural network on data, and it will learn to say, "I'm sure," or "I'm not sure," in a way that is actually useful.

This is a big deal because:

1. It's Faster: Calculating the "Logical Gap" with the Math Rulebook can be slow and expensive, especially for complex puzzles. The Neural Network just gives the answer in one quick step.
2. It's Flexible: The Math Rulebook relies on specific rules that might not work for every type of puzzle. The Neural Network learns from the data itself, so it can adapt to different types of noise or errors without needing a new rulebook.

In short: The paper shows that a "smart" computer program can learn to trust its own gut feeling about whether it's right or wrong, and that gut feeling is actually more accurate and useful than the traditional mathematical ruler scientists have been using for a long time.

Technical Summary

Technical Summary: Neural Network Decoder Confidence as a Learned Proxy for the Logical Gap

Problem Statement
In quantum error correction, decoders must infer the logical sector from measured syndromes. While Minimum-Weight Perfect Matching (MWPM) is a standard decoding algorithm, it also provides a "soft" confidence measure known as the complementary (or logical) gap. This gap, defined as the difference in weight between the minimum-weight correction in the predicted logical sector and the complementary sector, is used for post-selection: discarding low-confidence runs to reduce the logical error rate (LER).

Neural network (NN) decoders, particularly Graph Neural Networks (GNNs), have demonstrated superior hard-decision accuracy compared to MWPM. However, unlike MWPM, NN decoders do not inherently possess an interpretable confidence metric derived from an explicit optimization objective (such as a matching graph). It remains an open question whether the raw output (logit) of a trained NN can serve as a reliable, learned proxy for the logical gap, enabling effective post-selection without requiring the construction of competing logical sectors or explicit error models.

Methodology
The authors evaluate a pretrained GNN decoder (introduced in Ref. [1]) against a standard MWPM decoder on the rotated surface code under uniform circuit-level noise.

• Input and Architecture: The GNN takes a graph of detection events as input, where nodes are annotated with space-time coordinates and stabilizer types (X or Z). Edge weights are based on Euclidean distance. The network processes this through graph-convolution layers and a global pooling operation to output a single scalar logit ($z(s)$) for the logical observable.
• Confidence Metrics:
  • MWPM: The signed gap is defined as $g_{MWPM} = \omega(l_{wrong}) - \omega(l_{correct})$, where $\omega$ is the sum of edge weights ($\ln((1-p_e)/p_e)$).
  • GNN: The confidence score is the magnitude of the pre-sigmoid logit, $g_{GNN} = |z(s)|$. The sign is assigned retrospectively based on the known logical outcome for analysis (positive for correct, negative for failure).
• Evaluation: Both decoders are tested on the same sampled syndromes. The study analyzes three properties:
  1. Ranking: The ability to order shots by reliability via post-selection curves (LER vs. acceptance rate $\kappa$).
  2. Distribution: The shape of the signed confidence distributions.
  3. Calibration: How well the confidence magnitude predicts the conditional probability of logical failure, modeled by $P(\text{error}|g) = 1 / (1 + 10^{\alpha g/10})$, where $\alpha=1$ represents the ideal posterior log-likelihood ratio.

Key Contributions and Results

1. Superior Post-Selection Performance: At a fixed acceptance rate, post-selection based on the GNN logit yields a lower logical error rate than post-selection based on the MWPM gap. This holds across various code distances ($d=5, 7, 9$) and physical error rates. The GNN effectively retains more reliable shots without requiring a higher discard rate.
2. Gap-Like Distribution Structure: The signed confidence distribution of the GNN resembles the MWPM gap at low and intermediate values. However, in the high-confidence regime, the GNN assigns significantly higher confidence to correctly decoded shots without a corresponding increase in high-confidence failures. This asymmetry allows the GNN to push the acceptance threshold higher, driving the improved LER.
3. Quantitative Calibration: The GNN confidence magnitude follows the expected relation between confidence and failure probability more closely than the MWPM gap.
   • A pooled calibration fit across all simulated configurations yields a slope $\alpha = 0.93$ for the GNN, compared to $\alpha = 0.82$ for MWPM.
   • Since the ideal posterior log-likelihood ratio corresponds to $\alpha = 1$, the GNN provides a more faithful quantitative estimate of logical reliability, despite being trained only on syndromes and logical labels without explicit knowledge of the detector error model (DEM) weights.

Significance and Claims
The paper claims that a neural-network decoder trained solely on syndromes and logical labels can learn a soft information score that acts as a practical proxy for the logical gap. This finding is significant because:

• Generalizability: It suggests a route to gap-like post-selection in scenarios where MWPM-derived soft outputs are unavailable, computationally expensive, or difficult to define, such as in general qLDPC codes where the number of logical sectors grows exponentially ($4^k$).
• Efficiency: Unlike MWPM, which requires comparing minimum-weight corrections across competing sectors to compute a gap, the GNN provides a confidence score in a single forward pass.
• Learning Capability: The results demonstrate that NNs can learn not only to minimize hard-decision errors but also to internalize a quantitative confidence scale that approximates the ideal posterior log-odds, even without explicit constraints from the matching objective.

The authors maintain a modest stance, noting that while the GNN logit lacks the a priori interpretation of a logical gap, it empirically exhibits the necessary properties (ranking and calibration) to function effectively as one in post-selection protocols.