Curator's Take
This research challenges a fundamental assumption in quantum machine learning by showing that the widely-accepted trade-off between expressibility and trainability in quantum circuits largely breaks down when these circuits are embedded within hybrid classical-quantum neural networks. The finding that classical components can reshape the optimization landscape to decouple trainability from quantum circuit expressibility opens new possibilities for designing more powerful hybrid architectures without being constrained by traditional quantum circuit limitations. The introduction of multi-objective neural architecture search for joint classical-quantum optimization represents a significant methodological advance that could accelerate the development of practical quantum machine learning applications. This work suggests that the future of quantum ML may lie not in optimizing quantum circuits in isolation, but in co-designing classical and quantum components as integrated systems.
— Mark Eatherly
Summary
Hybrid quantum neural networks (HQNNs) integrate parameterized quantum circuits (PQCs) within classical networks, where the behavior of the underlying PQCs is often the primary focus of analysis. In this context, expressibility and trainability are widely used to characterize PQC's performance and are commonly assumed to exhibit a trade-off, where highly expressive circuits are more susceptible to barren plateaus. However, the validity of this relationship in HQNNs remains unclear. In this paper, we systematically analyze the expressibility--trainability relationship in HQNNs across varying circuit depths, qubit counts, entanglement topologies. We consider different training configurations, including pure PQCs, quantum-only training in hybrid setting, and full end-to-end training of hybrid models. Our results show that pure PQCs exhibit only a weak and regime-dependent trade-off, while hybrid architectures increasingly disrupt and can eliminate this relationship under full hybrid training. This indicates that classical components reshape the optimization landscape, decoupling trainability from PQC expressibility. We further propose a multi-objective neural architecture search (NAS) framework that jointly optimizes expressibility, trainability, and task performance over a combined classical--quantum design space, revealing different Pareto-optimal solutions under full end-to-end and quantum only training in hybrid setting. different trainability definitions. Our results suggest that hybridization is not just an implementation detail, but a defining factor in the performance of quantum machine learning models.