Training deep quantum neural networks.

Area of Science:

• Quantum Computing
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks are highly successful in classical computing.
• Developing quantum neural networks is essential for fully quantum learning tasks.
• Quantum technology advancements necessitate quantum analogues of classical computational models.

Purpose of the Study:

• To propose a truly quantum analogue of classical neurons.
• To construct quantum feedforward neural networks capable of universal quantum computation.
• To demonstrate efficient training methods for these quantum neural networks.

Main Methods:

• Introduced a novel quantum neuron design.
• Developed quantum feedforward neural networks utilizing these neurons.
• Employed fidelity as a cost function for efficient network training.
• Provided both classical and quantum implementations for training.

Main Results:

• Achieved efficient training with reduced memory requirements.
• Demonstrated that the number of qudits scales with network width, enabling deep network optimization.
• Benchmarked the proposal on the quantum task of learning an unknown unitary.
• Observed remarkable generalization behavior and robustness to noisy training data.

Conclusions:

• The proposed quantum neural networks offer a viable path for fully quantum learning.
• The efficient training method and resource scaling are significant advantages.
• The networks exhibit strong performance and robustness, crucial for practical quantum machine learning applications.