Speaker
Description
Quantum machine learning models typically rely on repeated quantum measurements and classical optimization during training, making the quantum-to-classical information-extraction process a fundamental bottleneck for scalability. We propose a framework for autonomous quantum machine learning based on Riemannian gradient optimization, in which, after encoding the training data, the optimization proceeds entirely within the quantum domain without intermediate measurements or classical feedback. We derive the corresponding Riemannian gradient update rules for a broad class of learning objectives and identify sufficient conditions under which the resulting training dynamics become fully autonomous. Within this regime, we establish an exponential convergence guarantee (in the training iterations) for sufficiently small step sizes and show that the trained model automatically exploits hidden symmetries present in the training data without requiring them to be specified $\textit{a priori}$. We further analyze the circuit complexity of the resulting trained models, showing that the circuit depth grows exponentially with the number of training iterations. Nevertheless, through quantum phase classification and quantum unitary learning, we demonstrate that effective learning can still be achieved at low circuit depths with only a few training samples. Our results establish autonomous quantum machine learning as an alternative learning paradigm that removes the classical optimization loop while retaining strong theoretical guarantees and strong practical performance.
| Topic | Quantum algorithms for machine learning applications |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |