Quantum-aware gradient descent for hybrid machine learning architectures

A viable approach for applying quantum computing to practical problems is hybrid quantum-classical machine learning (QML) systems. However, gradient discrepancies between quantum and classical components make it difficult to optimize such systems. For hybrid QML models, we suggest a new gradient descent technique called Quantum-Aware Gradient Descent (QAGD), which is intended to increase training stability and convergence. Our method addresses noise sensitivity and gradient scale mismatches by introducing layer-specific learning rate techniques and gradient normalization. The objective of this approach is to improve robustness and performance in noisy intermediate-scale quantum (NISQ) settings.