1
$\begingroup$

I'm reading the paper of Google's neural network decoder for surface code, and I'm confused about how to generate the measurement results for pre-training dataset.

If I understand it correctly, the dataset for pre-training stage was drawn from detector error models derived from correlation analysis of even subsets of experimental data. However, we can only sample the detection events instead of the raw measurements given the DEM, and the map between the detection events and measurements is not a bijection obviously. While the pre-training stage leveraged both the measurements and detection events, how was the measurements generated properly from the DEMs?


EDITED:

I realized that they did not use the measurements of data qubits at the final round in the model to avoid leaking labels. Given the initial state, then the map between the measurements and dets is truly a bijection.

$\endgroup$

1 Answer 1

1
$\begingroup$

For our purposes, we simply aggregated the detection events from the DEM sampler across the time axis (mod 2). This yields 100% redundant information, of course, but it seemed to help training performance either way, as we show in the ablation section.

$\endgroup$
1
  • $\begingroup$ Thanks for your reply! Does it means that you simply treat the initial stabilizer values as zeros? In my intuition, using different initial stabilizer values which means aggregating the detection events with an extra binary vector chosen randomly across different subsets of training dataset, will be more consistent with the actual experiment setup. $\endgroup$
    – Inm
    Nov 17, 2023 at 13:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.