Speaker
Description
Data re-uploading offers a compelling paradigm for near-term quantum machine learning. In this work, we present the full experimental realization of the originally proposed data re-uploading on a tunable integrated photonic processor. While the deep circuit requirements of multi-layer architectures present severe coherence-time bottlenecks for alternative quantum hardware, the integrated photonic platform naturally accommodates this depth. Crucially, it demonstrates how parameterizing linear optical networks can execute non-linear computations without requiring non-linear optical media. Within the Probably Approximately Correct (PAC) learning framework, we prove that our separated scheme maintains a finite Vapnik-Chervonenkis (VC) dimension scaling linearly with depth. Conversely, we show that for the examined worst-case configuration, the compressed variant popular in other implementations exhibits an infinite VC dimension, breaking formal generalization guarantees. Furthermore, Hessian analysis under these worst-case conditions reveals that the separated architecture ensures a vastly flatter optimization landscape, yielding a maximum Hessian eigenvalue (quantifying the lower bound of sharpness) of $\approx 0.23$ compared to $\approx 7.23\times10^{12}$ for the compressed model. These results establish our hardware-validated scheme as a highly trainable subroutine with robust worst-case guarantees for scalable optical quantum learning, and simultaneously offers a blueprint for energy-efficient classical optical non-linear computation.
| Topic | Quantum variational circuits and their applications |
|---|