Single-layer diffractive nets stall at 64.2% on MNIST: folding edge detection into one metasurface lifts accuracy to 80.7%
Synopsis
Single-layer diffractive neural networks hit an accuracy ceiling because their wavefront modulation is limited. Edge detection and classification are now combined on one nonlocal metasurface through spin multiplexing: the co-polarized output component performs momentum-space filtering for real-time edge detection, while the cross-polarized component is geometrically phase-modulated to classify. Crescent-shaped nanopillars couple quasi-bound states in the continuum with magnetic dipole resonances, reaching about 55% polarization conversion efficiency. On MNIST, this edge-enhanced architecture raises single-layer accuracy from 64.2% to 80.7%.
Interpretation
Edge detection and diffractive-network classification now share one metasurface, split by the light's spin components so contours are extracted before classification. Single-layer diffractive neural networks had faced accuracy bottlenecks from limited wavefront modulation, and combining edge detection with diffractive-network classification on a single chip remained a challenge. The co-polarized component of the output light performs real-time edge detection while the cross-polarized component performs classification, sequentially on the same platform.
The split comes from the metasurface's dispersion and phase response rather than from an extra electronic stage. The dispersion of the nonlocal Huygens metasurface lets the co-polarized component filter in momentum space, while the cross-polarized component receives geometric phase modulation. Quasi-bound states in the continuum and magnetic dipole resonances are coupled in crescent-shaped nanopillars, giving a measured polarization conversion efficiency of about 55%.
Edge enhancement cuts data redundancy, lowering the classification burden on the single-layer network. The single-layer network's classification accuracy on MNIST was 64.2%. With edge enhancement, the same single-layer network reaches 80.7% on that dataset.
Perspective
The demonstrated scope is all-optical edge detection plus classification on a single nonlocal metasurface platform, validated with a single-layer network on the MNIST handwritten-digit dataset, at about 55% polarization conversion efficiency. It addresses researchers and photonic device engineers working on integrated all-optical machine vision and intelligent photonic computing.
The reported accuracy gain is limited to a single-layer network on MNIST, with no results given for other image tasks or more complex scenes. Polarization conversion efficiency is about 55%, and where the remaining light goes and how that affects classification is not described in the material.