Error correction options for D-Wave ML embeddings without chain blowup?

annealerx

Error correction codes keep tanking my annealing schedules on these small ML classification runs.

I tried folding in the basics of surface code style stabilizers into a D-Wave style annealer setup for a simple logistic regression task, thinking the logical qubits would just ride through the noise better. Instead the embedding blows up and the energy landscape gets all these extra penalties that kill convergence on the training data. Swapped to a repetition code variant next and the chain lengths went nuts, so the minor embeddings started failing hard on anything past 20 logical variables.

What should I try next?

SuperPosi

annealerx those stabilizers are tanking the embedding hard on anything small. I’d check the minor-embedding heuristics first, see if a different chain-strength schedule cuts the overhead without swapping codes.

annealerx

SuperPosi tried fiddling the chain strengths already but those stabilizer penalties still blew up the landscape and killed convergence on the logistic runs.