| Abstract: |
Quantum machine learning (QML) has rapidly emerged as one of the most consequential frontiers of contemporary artificial intelligence research, promising computational advantages over purely classical approaches for high-dimensional learning, combinatorial optimisation, and structured-data inference. However, the practical deployment of fully fault-tolerant quantum machine learning remains constrained by limited qubit counts, decoherence, gate-error rates, and noisy intermediate-scale quantum (NISQ) hardware limitations. Hybrid quantum–classical (HQC) models have therefore become the dominant near-term paradigm, integrating parameterised quantum circuits (PQCs) with classical neural networks, optimisers, and pre/post-processing pipelines to deliver scalable learning on real-world data. This paper examines the architecture, performance, and scalability properties of HQC models through a critical evaluation of recent global research outputs up to 2025. The objectives are to assess the current state of HQC model design and to evaluate their scalability, accuracy, and resource efficiency relative to purely classical baselines. Adopting a descriptive secondary research design, the study synthesises authoritative datasets and benchmark studies from IBM Quantum, Google Quantum AI, Xanadu PennyLane, MIT-IBM Watson AI Lab, and arXiv-indexed peer-reviewed publications. Findings reveal that HQC variational classifiers achieve 87–94% accuracy on standard benchmark datasets at substantially lower parameter counts than classical baselines; quantum kernel methods deliver up to 23% accuracy improvement on structured-data tasks; and hybrid generative models reduce training-energy consumption by 18–34% on selected workloads. The discussion underscores the urgent need for standardised benchmarking, noise-aware training, and tight classical–quantum co-design. The paper concludes that hybrid quantum–classical learning is technically scalable, scientifically maturing, and a practical pathway toward production-grade quantum advantage in machine learning. |
