• Journal of the Korea Society of Computer and Information
• /
• v.28 no.4
• /
• pp.31-39
• /
• 2023

In this paper, we propose an Internal/External Knowledge Distillation (IEKD), which utilizes both external correlations between feature maps of heterogeneous models and internal correlations between feature maps of the same model for transferring knowledge from a teacher model to a student model. To achieve this, we transform feature maps into a sequence format and extract new feature maps suitable for knowledge distillation by considering internal and external correlations through a transformer. We can learn both internal and external correlations by distilling the extracted feature maps and improve the accuracy of the student model by utilizing the extracted feature maps with feature matching. To demonstrate the effectiveness of our proposed knowledge distillation method, we achieved 76.23% Top-1 image classification accuracy on the CIFAR-100 dataset with the "ResNet-32×4/VGG-8" teacher and student combination and outperformed the state-of-the-art KD methods.

• Earthquakes and Structures
• /
• v.20 no.4
• /
• pp.405-415
• /
• 2021

Following the dynamic property analysis and elaboration, linear response spectrum analysis (RSA) and response history analysis (RHA) were conducted on a representative hyperbolic cooling towers (HCT) in present study. The seismic responses in tower shell were illustrated in detail, including the internal force amplitude, modal contribution, influence from damping ratio, comparison of results got from RSA and RHA and especially the latitude distributions of internal forces. The results show that the eigenmodes could be classified in a new method into four types according to their mode shapes and only the lateral bending modes and vertical stretching modes are meaningful for horizontal and vertical earthquake correspondingly. The bending modes and seismic deformation display the same feature which is global lateral bending accompanied by minute circular flow displacement of section. This feature also decides the latitude distributions of internal forces as sine or cosine. Moreover, the following method is also proposed for approximate estimation of internal force amplitudes without time-consuming response history analysis: getting the response spectrums of the selected ground accelerations and then comparing values of response spectrums at the natural period of first lateral bending mode because it is always prime dominant for horizontal seismic responses.

• MALSORI
• /
• no.33_34
• /
• pp.71-89
• /
• 1997

In the framework of linear representation which allows for no internal structure within features, there is no way to represent nonlinear phonological phenomena such as complex segments. This paper shows how we carl solve some problems of the linear feature theory in relation to the hierarchical feature theory. The purpose of this paper is to explain lateral assimilation under hierarchical feature representation. Although arguments for the position of classes of distinctive features have been made the position of (lateral) remains the issue of debate. Sagey(1988) argues that the feature [lateral] is structurally dependent on the root node. In contrast Rice & Avery (1991) put the feature (lated) under the spontaneous voicing. I have discussed previous studies of feature hierarchy and I propose a revised model of feature representation. Within this model I have shown how well feature geometry describes lateralization as feature spreading.

• Journal of IKEEE
• /
• v.13 no.3
• /
• pp.18-23
• /
• 2009

FEATURE is a computational method to recognize functional and structural sites for automatic protein function prediction. By profiling physicochemical properties around residues, FEATURE can characterize and predict functional and structural sites in 3D protein structures in a high-throughput manner. Despite its effectiveness, it has been challenging to apply FEATURE to novel protein data due to limited customization support. To address this problem, we thoroughly analyze the internal modules of FEATURE and propose a methodology to customize FEATURE so that it can be used for new protein data for automatic functional annotations.

• Composites Research
• /
• v.17 no.1
• /
• pp.1-9
• /
• 2004

A new non-destructive evaluation method using tapping sound is proposed. This method, named Tapping Sound Analysis, is using the difference between tapping sound data of healthy structure and defective structure as the criteria of determination of internal defect of composite structure. For the characterization of tapping sound, a feature extraction method based on wavelet packet transform is proposed. And a feature index is defined for the decision of existence of internal defects. To prove the possibility of proposed method as a non-destructive evaluation method, experimental study is performed. The tapping sound data of healthy structure and defective structure are measured and compared based on the proposed decision method. The experimental results showed that the feature index is a good indicator for the determination of internal defects.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.5 no.3
• /
• pp.575-591
• /
• 2011

We present a novel active contour-based two-pass approach to extract smooth feature regions from a triangular mesh. In the first pass, an active contour formulated in level-set surfaces is devised to extract feature regions with rough boundaries. In the second pass, the rough boundary curve is smoothed by minimizing internal energy, which is derived from its curvature. The separation of the extraction and smoothing process enables us to extract feature regions with smooth boundaries from a triangular mesh without user's initial model. Furthermore, smooth feature curves can also be obtained by skeletonizing the smooth feature regions. We tested our algorithm on facial models and proved its excellence.

• International Journal of CAD/CAM
• /
• v.5 no.1
• /
• pp.39-47
• /
• 2005

Different CAx systems are being utilized throughout the product lifecycle due to the practical reasons in the supply chain and design processes. One of the major problems facing enterprises of today is how to share and exchange data among heterogeneous applications. Since different software applications use different terminologies, it is difficult to share and exchange the product data with internal and external partners. This paper presents a method to enhance the CAD model interoperability based on feature ontology. The feature ontology has been constructed based on the feature definition of modeling commands of CAD systems. A method for integration of semantic data has been proposed, implemented, and tested with two commercial CAD systems.

• Journal of Yeungnam Medical Science
• /
• v.29 no.2
• /
• pp.118-120
• /
• 2012

Loeffler's endocarditis involves progressive eosinophilic infiltration of the endocardium, which leads to apical thrombotic obliteration of the ventricle and endomyocardial fibrosis, that may finally represent a characteristic feature of restrictive cardiomyopathy. This paper presents a case of a 44-year-old male with symptoms of dyspnea and peripheral hypereosinophilia, who was diagnosed with early stage Loeffler's endocarditis via multicardiac imaging modalities.

• Journal of the Korea Society of Computer and Information
• /
• v.20 no.8
• /
• pp.35-43
• /
• 2015

Recently, internal information leaks is increasing rapidly by internal employees and authorized outsourcing personnel. In this paper, we propose a method to integrate internal control systems like system access control system and Digital Rights Managements and so on through expansion model of SIEM(Security Information Event Management system). this model performs a analysis step of security event link type and validation process. It develops unit scenarios to react illegal acts for personal information processing system and acts to bypass the internal security system through 5W1H view. It has a feature that derives systematic integration scenarios by integrating unit scenarios. we integrated internal control systems like access control system and Digital Rights Managements and so on through expansion model of Security Information Event Management system to defend leakage of internal information and customer information. We compared existing defense system with the case of the expansion model construction. It shows that expanding SIEM was more effectively.

• Journal of Korean Society for Geospatial Information Science
• /
• v.6 no.1 s.11
• /
• pp.27-34
• /
• 1998

A Unique Feature Identifier(UFID) is a way of referring to a feature, generally representing a tangible feature in the real world. In other words, a UFID uniquely identifies the related feature in the database and is normally used to link two or more databases together. This paper presents a UFID system aiming at the internal uses for National Geography Institute(NGI) as well as external uses for National Geographic Information System(NGIS) generally to link datasets together. The advantage of the proposed type of UFID lies in the meaningful nature of the identifier in providing a direct spatial index - administrative area and feature code. The checksum character proposed in this research is designed to remove any uncertainty about the number being corrupt. It will account lot digit transposition during manual input as well as corruption in transfer or processing.