Classifying single-qubit noise using machine learning
Travis L. Scholten, Yi-Kai Liu, Kevin Young, Robin Blume-Kohout
As quantum information processors (QIPs) grow more sophisticated, characterizing their behavior becomes harder. Larger QIPs have more properties that need to be characterized, and inventing a new quantum characterization, verification, and validation (QCVV) technique to probe a new property of interest requires time and effort. New approaches will be required to keep pace with the development of QIPs. Here, we investigate how machine learning (ML) algorithms can help automate the development of new QCVV techniques. A QCVV technique is, at heart, an analysis map from experimental data to an estimate of a property. ML algorithms are not, themselves, QCVV techniques. Instead, ML algorithms can create ("learn") new QCVV techniques by analyzing training examples. We identify the critical components of a "machine-learned QCVV" technique, and present a rubric for for using ML to create new QCVV techniques. To demonstrate this paradigm, we deploy several learning algorithms to create protocols for determining whether a qubit suffers from coherent or stochastic (incoherent) noise. Success depends on geometry specifically, whether the embedding of the noise classes into the dataset is consistent with the geometry of the ML algorithms hypothesis class. QCVV datasets generated by long circuits can reliably be separated by linear surfaces, but ones produced by short circuits (like the ones used for linear gate set tomography) may not be linearly separable. In these cases, feature engineering can enable linear classifiers to construct nonlinear QCVV techniques that classify noise instances correctly. When significant finite-sample noise is introduced in the data, we find that the support vector machine (SVM) algorithm continues to classify robustly.