📝 SummarXiv

Abstract

Effective training of Quantum Neural Networks (QNNs) is crucial in the Noisy Intermediate-Scale Quantum (NISQ) era, where noise accelerates the onset of barren plateaus (BPs) and limits scalability. This paper investigates how quantum noise impacts QNN trainability and demonstrates that careful selection of qubit measurement observables can mitigate these effects. We analyze PauliX, PauliY, PauliZ, and a customized Hermitian observable under both global (all-qubit measured) and local (single-qubit measured) cost functions. Our results show that with global cost function, PauliX and PauliY lead to flatter landscapes under noise, while PauliZ maintains training up to $8$ qubits before encountering BPs. The customized Hermitian observable proves most robust, enabling training up to $10$ qubits in noisy settings. For local cost function setting, PauliZ outperforms PauliX and PauliY, maintaining efficiency up to $10$ qubits. These findings highlight the importance of noise-aware observable selection, offering a practical strategy to improve QNN performance and advance quantum machine learning in noisy environments.