Towards Practical Quantum Neural Network Diagnostics with Neural Tangent Kernels

Knowing whether a Quantum Machine Learning model would perform well on a given dataset before training it can help to save critical resources. However, gathering a priori information about model performance (e.g., training speed, critical hyperparameters, or inference capabilities on unseen data) is a highly non-trivial task, in general. Recently, the Quantum Neural Tangent Kernel (QNTK) has been proposed as a powerful mathematical tool to describe the behavior of Quantum Neural Network (QNN) models. In this work, we propose a practical framework allowing to employ the QNTK for QNN performance diagnostics. More specifically, we show how a critical learning rate and a characteristic decay time for the average training error can be estimated from the spectrum of the QNTK evaluated at the initialization stage. We then show how a QNTK-based kernel formula can be used to analyze, up to a first-order approximation, the expected inference capabilities of the quantum model under study. We validate our proposed approach with extensive numerical simulations, using different QNN architectures and datasets. Our results demonstrate that QNTK diagnostics yields accurate approximations of QNN behavior for sufficiently deep circuits, can provide insights for shallow QNNs, and enables detecting - hence also addressing - potential shortcomings in model design.