Reproduce “Power of data in quantum machine learning” on OpenResearch

About this paper

Hsin-Yuan Huang, Michael Broughton, Masoud Mohseni, Ryan Babbush, Sergio Boixo, Hartmut Neven, Jarrod R. McClean

Researchers from Google Research quantitatively investigated the impact of training data on the competitiveness of classical machine learning against quantum models. They developed a framework using geometric difference to assess quantum prediction advantage and introduced "projected quantum kernels," which achieved over 20% accuracy gain on engineered datasets up to 30 qubits compared to classical baselines.

Key insights

• Classical machine learning models, given sufficient training data, can learn functions that are classically hard to compute for individual instances, often making them highly competitive with quantum models.
• The 'geometric difference' between classical and quantum kernels is a crucial function-independent metric: a small difference predicts no quantum advantage, while a large difference indicates potential.
• Standard quantum kernels often exhibit poor generalization due to the high dimensionality of Hilbert space, suggesting that projections (like in PQK) are necessary to create a meaningful separation from classical methods.

Summary by alphaXiv.