A belated update that about 2 months ago, I released another paper. This work, “Classifying Single-Qubit Noise Using Machine Learning”, was the result of a collaboration between myself, my PhD advisor (Robin Blume-Kohout), Yi-Kai Liu, and Kevin Young.
We considered a simple problem – determining whether the noise affecting a single qubit was coherent or stochastic – and looked at the performance of some simple algorithms (such as perceptrons and support vector machines) for this task. The key insight, in my opinion, is that a given experiment design – a set of circuits to be run on a quantum information processor – has an associated geometry resulting from performing those circuits on a noisy processor. Consequently, the accuracy of any classification technique will involve the interplay of the geometry of the separating surfaces in the hypothesis class of the algorithm and that “natural” geometry of the data.
For the experiment design we consider (gate set tomography), we show that under coherent and stochastic noise, the experiment design for linear gate set tomography is natural separated by a curved surface. This means algorithms whose hypothesis class contains only hyperplanes (such as perceptrons or linear support vector machines) will generally have low accuracy. We use feature engineering to embed the GST data sets into a new feature space where a natural separating surface is a hyperplane.
Overall, I’m very pleased with how this work turned out. It was my first foray into machine learning, and gave me a chance to play around with scikit-learn.
Do check it out! If you find any typos or other issues, please let me know.