• Kyungpook Mathematical Journal
• /
• v.22 no.2
• /
• pp.265-277
• /
• 1982

• ETRI Journal
• /
• v.40 no.4
• /
• pp.421-434
• /
• 2018

The application of deep neural networks (DNNs) to connect the world with cyber physical systems (CPSs) has attracted much attention. However, DNNs require a large amount of memory and computational cost, which hinders their use in the relatively low-end smart devices that are widely used in CPSs. In this paper, we aim to determine whether DNNs can be efficiently deployed and operated in low-end smart devices. To do this, we develop a method to reduce the memory requirement of DNNs and increase the inference speed, while maintaining the performance (for example, accuracy) close to the original level. The parameters of DNNs are decomposed using a hybrid of canonical polyadic-singular value decomposition, approximated using a tensor power method, and fine-tuned by performing iterative one-shot hybrid fine-tuning to recover from a decreased accuracy. In this study, we evaluate our method on frequently used networks. We also present results from extensive experiments on the effects of several fine-tuning methods, the importance of iterative fine-tuning, and decomposition techniques. We demonstrate the effectiveness of the proposed method by deploying compressed networks in smartphones.

• Journal of the Society of Naval Architects of Korea
• /
• v.33 no.3
• /
• pp.94-102
• /
• 1996

An interpretation for the additive and multiplicative decomposition theory of the deformation gradient tensor in finite deformation problems is presented. the conventional methods have not provided the additive deformation velocity gradient. Moreover the plastic deformation velocity gradients are not free from elastic deformations. In this paper, a modified multiplicative decomposition is introduced with the assumption of coaxial plastic deformation velocity gradient. This strategy well gives the additive deformation velocity gradient in which the plastic deformation velocity gradient is not affect4d by the elastic deformation.

• Bulletin of the Korean Mathematical Society
• /
• v.58 no.1
• /
• pp.253-267
• /
• 2021

We study linearly dependent subsets with prescribed cardinality s of a multiprojective space. If the set S is a circuit, there is an upper bound on the number of factors of the minimal multiprojective space containing S. B. Lovitz gave a sharp upper bound for this number. If S has higher dependency, this may be not true without strong assumptions (and we give examples and suitable assumptions). We describe the dependent subsets S with #S = 6.

• Bulletin of the Korean Mathematical Society
• /
• v.54 no.2
• /
• pp.463-484
• /
• 2017

We construct a weak Hopf algebra $wX_q(A_1)$ corresponding to non-standard quantum group $X_q(A_1)$. The PBW basis of $wX_q(A_1)$ is described and all the highest weight modules of $wX_q(A_1)$ are classified. Finally we give the Clebsch-Gordan decomposition of the tensor product of two highest weight modules of $wX_q(A_1)$.

• Korean Journal of Mathematics
• /
• v.30 no.1
• /
• pp.61-66
• /
• 2022

A new type of Riemannian manifold called semirecurrent manifold has been defined and some of its geometric properties are studied. Among others we show that the scalar curvature of semirecurrent manifold is constant and hence semirecurrent manifold is also concircularly recurrent. In addition, we show that the associated 1-form (resp. the associated vector field) of semirecurrent manifold is closed (resp. an eigenvector of its Ricci tensor). Furthermore, we prove that if a Riemannian product manifold is semirecurrent, then either one decomposition manifold is locally symmetric or the other decomposition manifold is a space of constant curvature.

• Investigative Magnetic Resonance Imaging
• /
• v.11 no.2
• /
• pp.110-118
• /
• 2007

Purpose: The objective of this work to construct eigenvalue maps that have information of magnitude of three primary diffusion directions using diffusion tensor images. Materials and Methods: To construct eigenvalue maps, we used a 3.0T MRI scanner. We also compared the Moore-Penrose pseudo-inverse matrix method and the SVD (single value decomposition) method to calculate magnitude of three primary diffusion directions. Eigenvalue maps were constructed by calculating of magnitude of three primary diffusion directions. We did investigate the relationship between eigenvalue maps and fractional anisotropy map. Results: Using Diffusion Tensor Images by diffusion tensor imaging sequence, we did construct eigenvalue maps of three primary diffusion directions. Comparison between eigenvalue maps and Fractional Anisotropy map shows what is difference of Fractional Anisotropy value in brain anatomy. Furthermore, through the simulation of variable eigenvalues, we confirmed changes of Fractional Anisotropy values by variable eigenvalues. And Fractional anisotropy was not determined by magnitude of each primary diffusion direction, but it was determined by combination of each primary diffusion direction. Conclusion: By construction of eigenvalue maps, we can confirm what is the reason of fractional anisotropy variation by measurement the magnitude of three primary diffusion directions on lesion of brain white matter, using eigenvalue maps and fractional anisotropy map.

• Smart Structures and Systems
• /
• v.13 no.2
• /
• pp.257-280
• /
• 2014

In this paper, a novel PARAllel FACtor (PARAFAC) decomposition based Blind Source Separation (BSS) algorithm is proposed for modal identification of structures equipped with tuned mass dampers. Tuned mass dampers (TMDs) are extremely effective vibration absorbers in tall flexible structures, but prone to get de-tuned due to accidental changes in structural properties, alteration in operating conditions, and incorrect design forecasts. Presence of closely spaced modes in structures coupled with TMDs renders output-only modal identification difficult. Over the last decade, second-order BSS algorithms have shown significant promise in the area of ambient modal identification. These methods employ joint diagonalization of covariance matrices of measurements to estimate the mixing matrix (mode shape coefficients) and sources (modal responses). Recently, PARAFAC BSS model has evolved as a powerful multi-linear algebra tool for decomposing an $n^{th}$ order tensor into a number of rank-1 tensors. This method is utilized in the context of modal identification in the present study. Covariance matrices of measurements at several lags are used to form a $3^{rd}$ order tensor and then PARAFAC decomposition is employed to obtain the desired number of components, comprising of modal responses and the mixing matrix. The strong uniqueness properties of PARAFAC models enable direct source separation with fine spectral resolution even in cases where the number of sensor observations is less compared to the number of target modes, i.e., the underdetermined case. This capability is exploited to separate closely spaced modes of the TMDs using partial measurements, and subsequently to estimate modal parameters. The proposed method is validated using extensive numerical studies comprising of multi-degree-of-freedom simulation models equipped with TMDs, as well as with an experimental set-up.

• Smart Structures and Systems
• /
• v.30 no.3
• /
• pp.221-237
• /
• 2022

This paper is concerned with nonlinear modeling and efficient numerical simulation of electrostrictive materials and structures. Two types of such materials are considered: relaxor ferroelectric ceramics and electrostrictive polymers. For ceramics, a geometrically linear formulation is developed, whereas polymers are studied in a geometrically nonlinear regime. In the paper, we focus on constitutive modeling first. For the reversible constitutive response under consideration, we introduce the augmented Helmholtz free energy, which is composed of a purely elastic part, a dielectric part and an augmentation term. For the elastic part, we involve an additive decomposition of the strain tensor into an elastic strain and an electrostrictive eigenstrain, which depends on the polarization of the material. In the geometrically nonlinear case, a corresponding multiplicative decomposition of the deformation gradient tensor replaces the additive strain decomposition used in the geometrically linear formulation. For the dielectric part, we first introduce the internal energy, to which a Legendre transformation is applied to compute the free energy. The augmentation term accounts for the contribution from vacuum to the energy. In our formulation, the augmented free energy depends not only on the strain and the electric field, but also on the polarization and an internal polarization; the latter two are internal variables. With the constitutive framework established, a Finite Element implementation is briefly discussed. We use high-order elements for the discretization of the independent variables, which include also the internal variables and, in case the material is assumed incompressible, the hydrostatic pressure, which is introduced as a Lagrange multiplier. The elements are implemented in the open source code Netgen/NGSolve. Finally, example problems are solved for both, relaxor ferroelectric ceramics and electrostrictive polymers. We focus on thin plate-type structures to show the efficiency of the numerical scheme and its applicability to thin electrostrictive structures.

• 한국HCI학회:학술대회논문집
• /
• 2009.02a
• /
• pp.197-202
• /
• 2009

Active Appearance Models(AAMs)은 얼굴인식, 얼굴추적, 표정인식 뿐만 아니라 눈동자 추적과 같은 분야에도 적용되어 좋은 성능을 보여 주었다. 보통 AAM 을 생성하기 위해서는 얼굴 영상과 얼굴의 특징을 나타내는 점으로 구성된 매쉬로 이루어 지는 트레이닝 셋이 필요하다. AAM fitting algorithm 은 학습한 얼굴과 유사한 얼굴을 Fitting 할 때에는 뛰어난 성능을 보이지만 조명에 의한 그림자 또는 액세서리에 의한 얼굴의 피부 가림과 같이 전체 얼굴이 잘 나타나지 않는 불완전한 영상의 Fitting 은 입력영상과 템플릿 영상간의 오차가 커지기 때문에 실패할 가능성이 매우 높다. 본 논문에서 우리는 AAMs 에서 사용되는 PCA를 Higher-order Singular Value Decomposition(HOSVD)로 대체하여 이 문제를 보완하는 강화된 AAM 을 제안한다. 제안된 AAM 에는 기존에 사용하던 고유벡터와 함께 HOSVD 를 통해 획득할 수 있는 Eigen-Modes 를 추가하여 사용한다. 또한 우리는 Yale Face Database를 이용한 평가를 통해 제안된 AAM 이 기존 AAM 보다 불완전한 영상에 효과적으로 대응하는 것을 보여준다.