Photonic quantum kernels outperform classical machine learning in real experiments

In a study just published in Nature Photonics, researchers from the CNR - Istituto di Fotonica e Nanotecnologie (CNR-IFN), Politecnico di Milano and the University of Vienna experimentally demonstrate a quantum-enhanced machine learning protocol that outperforms leading classical algorithms.

Using an integrated photonic processor, the team implemented a so-called “kernel method” based on quantum interference between indistinguishable photons. This approach enables binary classification of data with higher accuracy than classical methods such as Gaussian kernels and neural tangent kernels.

The experiment uses a laser-written photonic chip, where data are encoded into phase shifts and processed via quantum interference of two photons. The resulting distributions are then analyzed using a “support vector machine”, a well-known machine learning algorithm. Crucially, the method does not require entangled states or complex control—just quantum interference between single photons.

This collaboration shows that even with current noisy intermediate-scale quantum (NISQ) technologies, quantum effects can enhance real machine learning tasks. Remarkably, even when quantum interference is removed—by using distinguishable photons—the photonic hardware still outperforms standard classical kernels, suggesting an inherent computational advantage of optical processors.

The results offer a promising route for low-power, scalable quantum-enhanced machine learning and open new directions for hybrid photonic-classical computation.