• Journal of the Korean Vacuum Society
• /
• v.5 no.2
• /
• pp.99-106
• /
• 1996

Two types of regularization method (singular system and HMP approaches) for generating depth-concentration profiles from angle-resolved XPS data were evaluated. Both approaches showed qualitatively similar results although they employed different numerical algorithms. The application of the regularization method to simulated data demonhstrates its excellent utility for the complex depth profile system . It includes the stable restoration of depth-concentration profiles from the data with considerable random error and the self choice of smoothing parameter that is imperative for the successful application of the regularization method. The self choice of smoothing parameter is based on generalized cross-validation method which lets the data themselves choose the optimal value of the parameter.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.11 no.1
• /
• pp.45-52
• /
• 2016

This paper suggests a new method of finding regularization parameter for image restoration problems. Wiener filter requires priori information such that power spectrums of original image and noise. Constrained least squares restoration also requires knowledge of the noise level. If the prior information is not available, separate optimization functions for Tikhonov regularization parameter are suggested in the literature such as generalized cross validation and L-curve criterion. In this paper, self-regularization method that connects bias term of augmented linear system and smoothing term of Tikhonov regularization is introduced in the frequency domain and applied to the image restoration problems. Experimental results show the effectiveness of the proposed method.

• Analytical Science and Technology
• /
• v.8 no.4
• /
• pp.707-714
• /
• 1995

Two types of regularization method (singular system and HMP approaches) for generating depth-concentration profiles from angle-resolved XPS data were evaluated. Both approaches showed qualitatively similar results although they employed different numerical algorithms. The application of the regularization method to simulated data demonstrates its excellent utility for the complex depth profile system. It includes the stable restoration of the depth-concentration profiles from the data with considerable random error and the self choice of smoothing parameter that is imperative for the successful application of the regularization method. The self choice of smoothing parameter is based on generalized cross-validation method which lets the data themselves choose the optimal value of the parameter.

• Smart Structures and Systems
• /
• v.15 no.4
• /
• pp.1103-1120
• /
• 2015

A self-sensing magnetorheological (MR) damper with embedded piezoelectric force sensor has recently been devised to facilitate real-time close-looped control of structural vibration in a simple and reliable manner. The development and characterization of the self-sensing MR damper are presented based on experimental work, which demonstrates its reliable force sensing and controllable damping capabilities. With the use of experimental data acquired under harmonic loading, a nonparametric dynamic model is formulated to portray the nonlinear behaviors of the self-sensing MR damper based on NARX modeling and neural network techniques. The Bayesian regularization is adopted in the network training procedure to eschew overfitting problem and enhance generalization. Verification results indicate that the developed NARX network model accurately describes the forward dynamics of the self-sensing MR damper and has superior prediction performance and generalization capability over a Bouc-Wen parametric model.

• Journal of the Korea Institute of Military Science and Technology
• /
• v.24 no.4
• /
• pp.382-391
• /
• 2021

Unsupervised domain adaptation often gives impressive solutions to handle domain shift of data. Most of current approaches assume that unlabeled target data to train is abundant. This assumption is not always true in practices. To tackle this issue, we propose a general solution to solve the domain gap minimization problem without any target data. Our method consists of two regularization steps. The first step is a pixel regularization by arbitrary style transfer. Recently, some methods bring style transfer algorithms to domain adaptation and domain generalization process. They use style transfer algorithms to remove texture bias in source domain data. We also use style transfer algorithms for removing texture bias, but our method depends on neither domain adaptation nor domain generalization paradigm. The second regularization step is a feature regularization by feature alignment. Adding a feature alignment loss term to the model loss, the model learns domain invariant representation more efficiently. We evaluate our regularization methods from several experiments both on small dataset and large dataset. From the experiments, we show that our model can learn domain invariant representation as much as unsupervised domain adaptation methods.

• Journal of Electrical Engineering and Technology
• /
• v.13 no.4
• /
• pp.1724-1731
• /
• 2018

The conventional polynomial neural network (PNN) is a classical flexible neural structure and self-organizing network, however it is not free from the limitation of overfitting problem. In this study, we propose a space search-optimized polynomial neural network (ssPNN) structure to alleviate this problem. Ranking selection is realized by means of ranking selection-based performance index (RS_PI) which is combined with conventional performance index (PI) and coefficients based performance index (CPI) (viz. the sum of squared coefficient). Unlike the conventional PNN, L2-norm regularization method for estimating the polynomial coefficients is also used when designing the ssPNN. Furthermore, space search optimization (SSO) is exploited here to optimize the parameters of ssPNN (viz. the number of input variables, which variables will be selected as input variables, and the type of polynomial). Experimental results show that the proposed ranking selection-based polynomial neural network gives rise to better performance in comparison with the neuron fuzzy models reported in the literatures.

• Bulletin of the Korean Mathematical Society
• /
• v.52 no.6
• /
• pp.1945-1962
• /
• 2015

This paper is mainly concerned with the formation of delta standing wave for the scalar conservation law with a linear flux function involving discontinuous coefficients by using the self-similar viscosity vanishing method. More precisely, we use the self-similar viscosity to smooth out the discontinuous coefficient such that the existence of approximate viscous solutions to the delta standing wave for the Riemann problem is established and then the convergence to the delta standing wave solution is also obtained when the viscosity parameter tends to zero. In addition, the Riemann problem is also solved with the standard method and the instability of Riemann solutions with respect to the specific small perturbation of initial data is pointed out in some particular situations.

• Journal of the Korea Computer Graphics Society
• /
• v.6 no.1
• /
• pp.19-27
• /
• 2000

This paper proposes a non-self-intersecting multiresolution deformable model to extract and reconstruct three-dimensional boundaries of objects from volumetric data. Deformable models offer an attractive method for extracting and reconstructing the boundary surfaces. However, convensional deformable models have three limitations- sensitivity to model initialization, difficulties in dealing with severe object concavities, and model self-intersections. We address the initialization problem by multiresolution model representation, which progressively refines the deformable model based on multiresolution volumetric data in order to extract the boundaries of the objects in a coarse-to-fine fashion. The concavity problem is addressed by mesh size regularization, which matches its size to the unit voxel of the volumetric data. We solve the model self-intersection problem by including a non-self-intersecting force among the customary internal and external forces in the physics-based formulation. This paper presents results of applying our new deformable model to extracting a sphere surface with concavities from a computer-generated volume data and a brain cortical surface from a MR volume data.

• Journal of the Korean Institute of Electrical and Electronic Material Engineers
• /
• v.16 no.9
• /
• pp.801-806
• /
• 2003

In recent years, there has been large interest in the fabrication of the self organized nanoscale structures since not only their potential utilization in electronic, optoelectronic, and magnetic devices but also their fundamental interest such as uniformity and regularization. An attractive candidate of these materials is anodic porous alumina film(Al$_2$O$_3$) which is formed by the anodization of aluminum in an appropriate acid solution. In this study to fabricate the porous alumina film with very uniform and nearly parallel pores the anodization was carried out under constant voltage mode in 0.3M oxalic acid as an electrolyte. The hexagonally ordered arrays with a few $\mu\textrm{m}$ in size two-dimensional polycrystalline structure were obtained of which pore densities were 1.1${\times}$10$\^$10//$\textrm{cm}^2$.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.497-504
• /
• 2020

Constructing speaker embeddings in speaker verification is an important issue. In general, a self-attention mechanism has been applied for speaker embedding encoding. Previous studies focused on training the self-attention in a high-level layer, such as the last pooling layer. In this case, the effect of low-level layers is not well represented in the speaker embedding encoding. In this study, we propose Masked Cross Self-Attentive Encoding (MCSAE) using ResNet. It focuses on training the features of both high-level and low-level layers. Based on multi-layer aggregation, the output features of each residual layer are used for the MCSAE. In the MCSAE, the interdependence of each input features is trained by cross self-attention module. A random masking regularization module is also applied to prevent overfitting problem. The MCSAE enhances the weight of frames representing the speaker information. Then, the output features are concatenated and encoded in the speaker embedding. Therefore, a more informative speaker embedding is encoded by using the MCSAE. The experimental results showed an equal error rate of 2.63 % using the VoxCeleb1 evaluation dataset. It improved performance compared with the previous self-attentive encoding and state-of-the-art methods.