Quantifying the effect of quantum operations in machine learning by comparison to classically simulable variants

RWTH Publications (RWTH Aachen)

Abstract

Quantum machine learning (QML) investigates whether quantum-mechanical representations and computations can improve machine-learning tasks. However, it is often unclear which observed performance differences are genuinely attributable to specifically quantum features rather than to architectural differences that can be implemented classically or benchmarking effects. This thesis studies the role of entanglement in variational QML models through a controlled ablation or removal framework. The work introduces partially and fully separable variants of the Circuit-centric Classifier (CCC) and IQP-Kernel Classifier, constructed by systematically reducing entangling interactions while keeping the remaining training and evaluation pipeline fixed. The models are evaluated on multiple synthetic and structured benchmark families, including but not limited to linearly separable, hidden-manifold, and hyperplane-based datasets. The results show that partially entangling variants often remain competitive with the original models, whereas fully separable variants typically exhibit larger performance degradation. The experiments are consistent with entanglement contributing meaningfully to performance in the studied models, while also suggesting that sparse entangling connectivity may already retain much of the observed benefit of the original architectures in this benchmark setting. The IQP-kernel experiments scaled prohibitively in higher-dimensional settings, exceeding the available computational budget.

Authors 1

1. Víctor Mauricio Gutiérrez Luque corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen