Do quantum models make good generative learners?

Thesis

Utilizing quantum computers as models for generative machine learning became an influential idea around 2019, but with little evidence to support and understand this intuition. In this thesis, we present a collection of evidence that addresses the question: "Do quantum models make good generative learners?". We first introduce a novel definition and framework for evaluating quantum generative models as good, and use it to better understand the learning capabilities of the state-of-the-art Quantum Circuit Born Machine (QCBM). We then present a new kind of quantum generative model composed of parameterized repeat-until-success circuits with mid-circuit measurements - different to the QCBM with respect to the presence of non-linearity in its state evolution. We show the effectiveness and limitations that result from this particular type of non-linearity, and that it additionally can be modified into a useful method for evaluating more advanced quantum hardware. Overall, we observe that at small-scales and with enough resources, quantum models do make good generative learners. However, for them to be practically useful with large amounts of data or be intentionally used over a classical network, we need to create methods to efficiently train at large scale and to understand where the inductive biases of these models are directly aligned with the data.

Authors: Gili, K

• DOI: 10.5287/ora-erpryxaok
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; quantum computing
• Subjects: Machine Learning; Quantum Computing