We propose a simple two-step approach for speeding up convolution layers within large convolutional neural networks based on tensor decomposition and discriminative fine-tuning. Given a layer, we use non-linear least squares to compute a low-rank CP-decomposition of the 4D convolution kernel tensor into a sum of a small number of rank-one tensors. At the second step, this decomposition is used to replace the original convolutional layer with a sequence of four convolutional layers with small kernels. After such replacement, the entire network is fine-tuned on the training data using standard backpropagation process. We evaluate this approach on two CNNs and show that it is competitive with previous approaches, leading to higher obtained CPU speedups at the cost of lower accuracy drops for the smaller of the two networks. Thus, for the 36-class character classification CNN, our approach obtains a 8.5x CPU speedup of the whole network with only minor accuracy drop (1% from 91% to 90%). For the standard ImageNet architecture (AlexNet), the approach speeds up the second convolution layer by a factor of 4x at the cost of 1% increase of the overall top-5 classification error.
We demonstrate that variable angle spectroscopic ellipsometry (SE) is capable of determining both ordinary and extraordinary dielectric functions (DFs) in the transparent region of group III nitride wurtzite films even if the optical axis is oriented normal to the surface plane. In contrast to the so far used prism coupling technique, the SE method is neither restricted to layers deposited on lower refractive index substrates nor to available laser wavelengths. Application to AlN and GaN films grown on 6H-SiC substrates yields experimental data which can be represented in simple analytical form. The difference in energy dispersion for ordinary and extraordinary DFs is related to the effective optical band gap depending on the light polarization. Extrapolation of data to lower photon energies allows estimation of the ordinary and extraordinary high-frequency dielectric constants.
The influence of dislocations on electron transport properties of undoped InN thin films grown by molecular-beam epitaxy on AlN͑0001͒ pseudosubstrates is reported. The microstructure and the electron transport in InN͑0001͒ films of varying thickness were analyzed by transmission electron microscopy and variable temperature Hall-effect measurements. It was found that crystal defects have strong effects on the electron concentration and mobility of the carriers in the films. In particular, the combined analysis of microscopy and Hall data showed a direct dependence between free carrier and dislocation densities in InN. It was demonstrated that threading dislocations are active suppliers of the electrons and an exponential decay of their density with the thickness implies the corresponding decay in the carrier density. The analysis of the electron transport yields also a temperature-independent carrier concentration, which indicates degenerate donor levels in the narrow band-gap InN material. The relative insensitivity of the mobility with respect to the temperature suggests that a temperature-independent dislocation strain field scattering dominates over ionized impurity/defect and phonon scattering causing the increase of the mobility with rising layer thickness due to the reducing dislocation density. Room temperature mobilities in excess of 1500 cm 2 V −1 s −1 were obtained for ϳ800 nm thick InN layers with the dislocation densities of ϳ3 ϫ 10 9 cm −2 .
Monosized (∼4 nm) diamond nanoparticles arranged on substrate surfaces are exciting candidates for single-photon sources and nucleation sites for ultrathin nanocrystalline diamond film growth. The most commonly used technique to obtain substrate-supported diamond nanoparticles is electrostatic self-assembly seeding using nanodiamond colloidal suspensions. Currently, monodisperse nanodiamond colloids, which have a narrow distribution of particle sizes centering on the core particle size (∼4 nm), are available for the seeding technique on different substrate materials such as Si, SiO2, Cu, and AlN. However, the self-assembled nanoparticles tend to form small (typically a few tens of nanometers or even larger) aggregates on all of those substrate materials. In this study, this major weakness of self-assembled diamond nanoparticles was solved by modifying the salt concentration of nanodiamond colloidal suspensions. Several salt concentrations of colloidal suspensions were prepared using potassium chloride as an inserted electrolyte and were examined with respect to seeding on SiO2 surfaces. The colloidal suspensions and the seeded surfaces were characterized by dynamic light scattering and atomic force microscopy, respectively. Also, the interaction energies between diamond nanoparticles in each of the examined colloidal suspensions were compared on the basis of the Derjaguin-Landau-Verwey-Overbeek (DLVO) theory. From these investigations, it became clear that the appropriate salt concentration suppresses the formation of small aggregates during the seeding process owing to the modified electrostatic repulsive interaction between nanoparticles. Finally, monosized (<10 nm) individual diamond nanoparticles arranged on SiO2 surfaces have been successfully obtained.
Articles you may be interested inImpact of varying buffer thickness generated strain and threading dislocations on the formation of plasma assisted MBE grown ultra-thin AlGaN/GaN heterostructure on silicon AIP Advances 5, 057149 (2015); 10.1063/1.4921757 Improved electron mobility in InSb epilayers and quantum wells on off-axis Ge (001) substratesThe strain-relaxation phenomena and the formation of a dislocation network in 2H-InN epilayers during molecular beam epitaxy are reported. Plastic and elastic strain relaxations were studied by reflection high-energy electron diffraction, transmission electron microscopy, and high resolution x-ray diffraction. Characterization of the surface properties has been performed using atomic force microscopy and photoelectron spectroscopy. In the framework of the growth model the following stages of the strain relief have been proposed: plastic relaxation of strain by the introduction of geometric misfit dislocations, elastic strain relief during island growth, formation of threading dislocations induced by the coalescence of the islands, and relaxation of elastic strain by the introduction of secondary misfit dislocations. The model emphasizes the determining role of the coalescence process in the formation of a dislocation network in heteroepitaxially grown 2H-InN. Edge-type threading dislocations and dislocations of mixed character have been found to be dominating defects in the wurtzite InN layers. It has been shown that the threading dislocation density decreases exponentially during the film growth due to recombination and, hence, annihilation of dislocations, reaching ϳ10 9 cm −2 for ϳ2200 nm thick InN films.
A model for the influence of different contributions to the high electron concentration in dependence on the film thickness of state-of-the-art InN layers grown by molecular-beam epitaxy is proposed. Surface accumulation has a crucial influence for InN layers <300nm and superimposes the background concentration. For air-exposed InN, it can be assigned to a surface near doping by oxygen. For InN layers in the micron range the density of dislocations is the major doping mechanism. Finally, point defects such as vacancies and impurities have minor influence and would dominate the free electron concentration only for InN >10μm.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.