Speaker
Description
New decoders for quantum error correction are almost always judged by how much they beat a standard classical decoder, BP-OSD. We show this verdict can hinge on a single, easily overlooked setting of that baseline, with no safe default. The setting (the belief-propagation min-sum scaling factor) can change the baseline's error rate more than a hundredfold, and the value best for one kind of noise is the worst for another, flipping between idealized (code-capacity) and realistic (circuit-level) noise, a reversal we confirm across two code families and two code sizes. We give a simple recipe for tuning the baseline in the regime actually being reported, plus a hardware-independent cost measure so accuracy and decoding cost can be weighed together. As an illustration, a small neural-network decoder we train looks competitive with the baseline under one tuning but trails it by 30 to 65 times under a fair one. The takeaway: tune your baseline for the noise you report, and say how, before claiming to beat it.
| Topic | Machine learning techniques for experimental quantum information science |
|---|