Reproduce “Enhancing the Scalability of Classical Surrogates for Real-World Quantum Machine Learning Applications” on OpenResearch

About this paper

Philip Anton Hernicht, Alona Sakhnenko, Corey O'Meara, Giorgio Cortiana, Jeanette Miriam Lorenz

This work introduces an optimized pipeline for generating classical surrogates of trained quantum machine learning models, effectively reducing the computational demands from exponential to linear scaling. This advancement enables the conversion of larger quantum models, such as a 9-qubit, 2-layer model on a standard laptop, thus facilitating the practical deployment of QML in industrial settings and supporting empirical investigations into quantum advantage.

Key insights

• Variational reuploading quantum models can be represented as truncated Fourier series, and their coefficients can be efficiently approximated classically.
• The primary limitation of prior surrogation methods was the exponential memory scaling associated with generating a comprehensive input grid and the full Fourier spectrum.
• Significant redundancies exist in both the input sampling and frequency spectrum, which can be exploited to reduce computational resources for surrogate generation.

Summary by alphaXiv.