Shadows of quantum machine learning

Area of Science:

• Quantum Computing
• Machine Learning
• Computational Complexity

Background:

• Quantum machine learning (QML) offers computational advantages but requires quantum hardware for evaluation.
• Evaluating trained QML models on new data necessitates access to quantum computers, limiting practical use.

Purpose of the Study:

• To introduce a novel class of QML models trainable with quantum resources but deployable classically.
• To enable practical QML applications by decoupling training from evaluation.

Main Methods:

• Developed a training methodology resulting in a 'shadow model' for classical deployment.
• Proved universality for classically-deployed QML.
• Analyzed learning capacities and compared them to fully quantum and classical models.

Main Results:

• The proposed models are universal for classically-deployed QML.
• These models exhibit restricted learning capacities compared to fully quantum models.
• A provable learning advantage over classical learners is achieved under standard complexity assumptions.

Conclusions:

• Quantum machine learning can offer advantages even when quantum computers are used solely for training.
• Classical deployment of QML models broadens their applicability in real-world scenarios.
• This research facilitates the integration of QML into various practical contexts by overcoming hardware limitations for evaluation.