February 2020
Spotlight Summary by David Powell
Neuromorphic metasurface
The potential of metasurfaces to replace conventional optical elements is widely appreciated. Here, Wu et al. apply them to a completely different field: machine learning. The team from University of Wisconsin considers the problem of recognizing handwritten numbers, designing and modelling "neuromorphic metasurfaces" which can achieve 80–90% recognition accuracy. The structures consist of several layers of nano-structured dielectric metasurfaces, where a single geometric parameter is varied across the surface to achieve phase control. The optical transfer function is modelled analytically, enabling it to be incorporated into existing machine learning software packages. The considered structure is one dimensional, a limitation largely imposed by the computational burden, which may be overcome with high-performance computing systems. Perhaps most surprisingly, the results are achieved using a purely linear metasurface, whereas modern techniques such as deep learning typically apply a nonlinear activation function to each layer. Achieving a strong nonlinear response within a thin metasurface remains challenging. However, great progress is being made by several research groups. Metasurfaces are one of many structures proposed under the emerging field of neuromorphic photonics, and it will be interesting to see whether their advantages of planar fabrication and high information density will make them a compelling platform.
You must log in to add comments.
Add Comment
You must log in to add comments.
Article Information
Neuromorphic metasurface
Zhicheng Wu, Ming Zhou, Erfan Khoram, Boyuan Liu, and Zongfu Yu
Photon. Res. 8(1) 46-50 (2020) View: Abstract | HTML | PDF