Quantum machine learning beyond kernel methods

Machine learning algorithms based on parametrized quantum circuits are a prime candidate for near-term applications on noisy quantum computers. Yet, our understanding of how these quantum machine learning models compare, both mutually and to classical models, remains limited. Previous works achieved important steps in this direction by showing a close connection between some of these quantum models and kernel methods, well-studied in classical machine learning. In this work, we identify the first unifying framework that captures all standard models based on parametrized quantum circuits: that of linear quantum models. In particular, we show how data re-uploading circuits, a generalization of linear models, can be efficiently mapped into equivalent linear quantum models. Going further, we also consider the experimentally-relevant resource requirements of these models in terms of qubit number and data-sample efficiency, i.e., amount of data needed to learn. We establish learning separations demonstrating that linear quantum models must utilize exponentially more qubits than data re-uploading models in order to solve certain learning tasks, while kernel methods additionally require exponentially many more data points. Our results constitute significant strides towards a more comprehensive theory of quantum machine learning models as well as provide guidelines on which models may be better suited from experimental perspectives.