This paper experimentally demonstrates an optical neural network achieving 99% accuracy on handwritten-digit classification using -3.1 detected photons per weight multiplication and -90% accuracy using -0.66 photons per weight multiplication. The sub-photon-per-multiplication is enabled by noise reduction from accumulating scalar multiplications in dot-product sums, applicable to various optical-neural-network architectures. This work highlights the potential of optical neural networks for accurate results with extremely low optical energies.

Tianyu Wang, Shi-Yuan Ma, Logan G. Wright, Tatsuhiro Onodera, Brian C. Richard, Peter L. McMahon