Robust deep learning object recognition models rely on low frequency information in natural images

Machine learning models have difficulty generalizing to data outside of the distribution they were trained on. In particular, vision models are usually vulnerable to adversarial attacks or common corruptions, to which the human visual system is robust. Recent studies have found that regularizing machine learning models to favor brain-like representations can improve model robustness, but it is unclear why. We hypothesize that the increased model robustness is partly due to the low spatial frequency preference inherited from the neural representation. We tested this simple hypothesis with several frequency-oriented analyses, including the design and use of hybrid images to probe model frequency sensitivity directly. We also examined many other publicly available robust models that were trained on adversarial images or with data augmentation, and found that all these robust models showed a greater preference to low spatial frequency information. We show that preprocessing by blurring can serve as a defense mechanism against both adversarial attacks and common corruptions, further confirming our hypothesis and demonstrating the utility of low spatial frequency information in robust object recognition.

Manipulation of Magnetic Dipole Emission from Eu3+ with Mie-Resonant Dielectric Metasurfaces

Mie-resonant high-index dielectric nanoparticles and metasurfaces have been suggested as a viable platform for enhancing both electric and magnetic dipole transitions of fluorescent emitters. While the enhancement of the electric dipole transitions by such dielectric nanoparticles has been demonstrated experimentally, the case of magnetic-dipole transitions remains largely unexplored. Here, we study the enhancement of spontaneous emission of Eu³⁺ ions, featuring both electric and magnetic-dominated dipole transitions, by dielectric metasurfaces composed of Mie-resonant silicon nanocylinders. By coating the metasurfaces with a layer of an Eu³⁺ doped polymer, we observe an enhancement of the Eu³⁺ emission associated with the electric (at 610 nm) and magnetic-dominated (at 590 nm) dipole transitions. The enhancement factor depends systematically on the spectral proximity of the atomic transitions to the Mie resonances as well as their multipolar order, both controlled by the nanocylinder size. Importantly, the branching ratio of emission via the electric or magnetic transition channel can be modified by carefully designing the metasurface, where the magnetic dipole transition is enhanced more than the electric transition for cylinders with radii of about 130 nm. We confirm our observations by numerical simulations based on the reciprocity principle. Our results open new opportunities for bright nanoscale light sources based on magnetic transitions.

Formation of nanocrystalline graphene on germanium

Graphitization of a polymer layer provides a convenient route to synthesize nanocrystalline graphene on dielectric surfaces. The transparent and conducting wafer scale material is of interest as a membrane and a coating, and for the generation and detection of light, or strain sensing. In this work, we study the formation of nanocrystalline graphene on germanium, a surface which promotes the CVD synthesis of monocrystalline graphene. The surprising result that we obtained through graphitization is the formation of cavities in germanium, over which nanocrystalline graphene is suspended. Depending on the crystallographic orientation of the germanium surface, either trenches in (110)-Ge or pits in (111)-Ge are formed, and their dimensions depend on the graphitization temperature. Using Raman spatial imaging, we can show that nanocrystalline graphene is formed across the entire wafer in spite of the cavity formation. Interestingly, the Raman intensity is suppressed when the material is supported by germanium and is enhanced when the material is suspended. Through simulations, we can show that these effects are induced by the high refractive index of germanium and by interferences of the light field depending on the spacing between graphene and germanium. Using atomic force and scanning electron microscopy, we determined that ripples in the suspended material are induced by the mismatch of thermal expansion coefficients. Our results provide a new route to lithography-free fabrication of suspended membranes.

Active tuning of all-dielectric metasurfaces

All-dielectric metasurfaces provide a powerful platform for highly efficient flat optical devices, owing to their strong electric and magnetic dipolar response accompanied by negligible losses at near-infrared frequencies. Here we experimentally demonstrate dynamic tuning of electric and magnetic resonances in all-dielectric silicon nanodisk metasurfaces in the telecom spectral range based on the temperature-dependent refractive-index change of a nematic liquid crystal. We achieve a maximum resonance tuning range of 40 nm and a pronounced change in the transmittance intensity up to a factor of 5. Strongly different tuning rates are observed for the electric and the magnetic response, which allows for dynamically adjusting the spectral mode separation. Furthermore, we experimentally investigate the influence of the anisotropic (temperature-dependent) dielectric environment provided by the liquid crystal on both the electric and magnetic resonances. We demonstrate that the phase transition of the liquid crystal from its nematic to its isotropic phase can be used to break the symmetry of the optical metasurface response. As such, our approach allows for spectral tuning of electric and magnetic resonances of all-dielectric metasurfaces as well as switching of the anisotropy of the optical response of the device.