Quantum Machine Learning (QML) aims to leverage the principles of quantum mechanics to speed up the process of solving machine learning problems or improve the quality of solutions. Among these principles, entanglement with an auxiliary system was shown to increase the quality of QML models in applications such as supervised learning. Recent works focus on the information that can be extracted from entangled training samples and their effect on the approximation error of the trained model. However, results on the trainability of QML models show that the training process itself is affected by various properties of the supervised learning task. These properties include the circuit structure of the QML model, the used cost function, and noise on the quantum computer. To evaluate the applicability of entanglement in supervised learning, we augment these results by investigating the effect of highly entangled training data on the model's trainability. In this work, we show that for highly expressive models, i.e., models capable of expressing a large number of candidate solutions, the possible improvement of loss function values in constrained neighborhoods during optimization is severely limited when maximally entangled states are employed for training. Furthermore, we support this finding experimentally by simulating training with Parameterized Quantum Circuits (PQCs). Our findings show that as the expressivity of the PQC increases, it becomes more susceptible to loss concentration induced by entangled training data. Lastly, our experiments evaluate the efficacy of non-maximal entanglement in the training samples and highlight the fundamental role of entanglement entropy as a predictor for the trainability.
Mandl et al. (Fri,) studied this question.