187037 research outputs found
Sort by
Method and Apparatus for Targeted Data Poisoning Using Probabilistic Decision Boundaries
No full text확률적 결정 경계를 사용한 표적화된 데이터 중독 방법 및 장치가 제시된다. 본 발명에서 제안하는 확률적 결정 경계를 사용한 표적화된 데이터 중독 장치는 확률적 결정 경계를 생성하기 위해 표적 데이터 주변의 데이터를 이용하여 대체 모델을 학습하고 저장하는 복수의 대체 모델 학습부, 상기 복수의 대체 모델 학습부로부터 학습된 데이터를 입력 받아 결과를 취합하여 표적 데이터 및 표적 데이터와 다른 라벨을 가질 확률을 계산하는 확률적 결정 경계 생성부, 현재 가지고 있는 데이터 이외에 실제 데이터와 유사한 데이터 후보를 생성하는 후보 생성부 및 상기 후보 생성부에서 생성된 데이터 후보 중 상기 확률적 결정 경계 생성부에서 계산된 확률을 이용하여 거짓 데이터를 생성하는 거짓 데이터 생성부를 포함한다
정보 출력 장치
No full textThis application relates to an information output apparatus including at least one information output unit. The information output unit includes a coil unit connected to a power source to allow an electric current to flow therethrough and a base unit formed to accommodate the coil unit. The information output unit also includes an expression unit formed and arranged to allow a user to sense the expression unit and a driving unit arranged in the base unit to be separated from the coil unit. The driving unit is adjacent to the coil unit to be driven by the electric current flowing through the coil unit to perform an angular or rotational movement so as to move the expression unit in a first direction towards the coil unit and an opposite direction
MANUFACTURING METHOD OF ZIRCONIUM NITRIDE NITROGEN CARRIER FOR AMMONIA SYNTHESIS AND ZIRCONIUM NITRIDE NITROGEN CARRIER FOR AMMONIA SYNTHESIS MANUFACTURED THEREBY
No full text본 발명은 매체순환 공정(chemical looping)에 의한 암모니아 생산에 사용되는 질소전달체에 관한 것으로, 요소 유리 경로(Urea-glass route, UGR)라는 이온 결합 구조체를 기반의 암모니아 생산용 질화지르코늄 질소전달체 제조방법 및 이에 의해 제조된 암모니아 생산용 질화지르코늄 질소전달체에 관한 것이다
타가토스 생산용 조성물 및 이를 이용한 타가토스 제조 방법
No full textThe present disclosure relates to a composition for tagatose production and a method for preparing tagatose using the same
긴 꼬리문제 상황에서 반복적인 메타머 기반 학습
No full text학위논문(박사) - 한국과학기술원 : 뇌인지공학프로그램, 2025.2,[iv, 41 p. :]The long-tail problem addresses the issue of extreme data imbalance, reflecting the distribution of real-world data collection. However, previous studies have focused on separate learning representations to enhance the performance of tail classes with insufficient data, without considering the human learning process. This approach leads to a decline in classification and generalization performance in situations with extremely limited data. In this paper, we propose a metameric recurrent training(MRT) inspired by human learning. It is based on metamerism, a phenomenon that the ventral stream of the human visual cortex exhibits similar neural representations for contextually-related visual stimuli. The proposed framework registers the latent representations of the training data in the forward prediction process, generates auxiliary samples matching those latent representations, and repeats this cycle. We confirmed improvements in classification and generalization performance and conducted a model analysis based on the information bottleneck theory. We verified whether the generated metamers guide the model to human-like learning through mTurk tasks and confirmed the applicability to real-world data.한국과학기술원 :뇌인지공학프로그램
빛에 반응하는 블록공중합체 입자 개발
No full text학위논문(박사) - 한국과학기술원 : 생명화학공학과, 2025.2,[viii, 152 p. :]Smart materials have long been a subject of great interest due to their ability to undergo reversible form changes in response to external stimuli. Especially, responsive particle system that exhibit shape-dependent changes in their optical and rheological properties have applications in many domains including displays, smart windows, films, and sensors. Among several stimuli, light has gained significant interest due to its remote controllability, potential wavelength-dependent effects, ability to be locally targeted, and rapid reaction capabilities. Consequently, there is a growing demand to create tiny particles that can alter their structure and characteristics in reaction to light.
An emulsion approach based on the self-assembly of block copolymers within an oil-in-water emulsion allows for the simple production of particles with diameters ranging from hundreds of nanometers to tens of micrometers. Furthermore, by controlling various thermodynamic and kinetic parameters, it is possible to generate particles with diverse forms and interior morphologies. For instance, the interfacial energy between polymer-polymer and polymer-continuous phases, as well as the interaction between polymer-solvent, can significantly impact the self-assembly behavior of block copolymers and the morphology of the resulting microparticles. This dissertation introduces a new system that incorporates reversible photoisomerizable molecules into block copolymer particles created by the oil-in-water emulsion process. This technique allows for the reversible modification of the shape and optical properties of the block copolymer particles in response to varying wavelengths of light.
In this study, we introduce a novel light-responsive block copolymer (BCP) particle system, capable of exhibiting rapid and reversible transformations in both shape and color upon exposure to light, marking a significant advancement in the development of programmable smart materials. Utilizing a series of spiropyran-based surfactants with varying alkyl spacer lengths, from hexyl to ethyl decanoate, we have successfully developed photoactive, shape-changing particles. In addition, the integration of spiropyran photoacid (SPPA) molecules into the BCP system enables dynamic morphological changes in response to 420 nm light irradiation. Furthermore, we develop a photoreactive hydrazone-based additive that enables particles to undergo shape and internal structural changes through photoinduced chemical reactions. These light-responsive block copolymer microparticles developed in this study are anticipated to serve as intelligent particles with versatile applications.한국과학기술원 :생명화학공학과
대규모 실내 환경에서 Wi-Fi 및 자기 신호의 상호보완적 융합을 이용한 그래프 기반 SLAM
No full text학위논문(박사) - 한국과학기술원 : 건설및환경공학과, 2025.2,[viii, 66 p. :]This study presents an efficient algorithm for user localization utilizing only the sensors embedded in a smartphone. Unlike conventional methods that require external devices or complex configurations, this approach leverages built-in sensors such as accelerometers, gyroscopes, Wi-Fi, and magnetometers to achieve practical and scalable localization. By excluding cameras, this research minimizes privacy concerns while enabling location estimation solely through a smartphone. Traditional fingerprinting-based Wi-Fi localization methods often face challenges related to environmental changes and the labor-intensive nature of data collection and maintenance. To address these issues, the proposed algorithm utilizes real-time Wi-Fi data, allowing localization without prior knowledge of Wi-Fi access point (AP) locations. Key inputs such as the number of APs, unique MAC addresses, and RSSI values enable accurate and reliable location estimation without the limitations of fingerprinting approaches.
Moreover, this study introduces a novel approach to enhance localization accuracy by complementarily integrating Wi-Fi and magnetic field data. While Wi-Fi excels in providing global localization capabilities, it suffers from reduced accuracy due to signal noise. Magnetic field data, on the other hand, offers high precision in local matching but can result in incorrect matches due to ambiguity. To mitigate these limitations, a path loss model is applied to restrict the search area during magnetic-based matching, effectively reducing erroneous constraints. Additionally, a filtering technique using Wi-Fi signal strength improves the reliability of magnetic field data, enabling precise location estimation in complex indoor environments.
The research further addresses the challenges of sparse data and noise by proposing a robust localization algorithm. Using CS techniques during the regression process, the algorithm ensures effective restoration even in environments with limited data, reducing dependency on training datasets or specific parameters. Extensive experiments were conducted in large-scale indoor settings, demonstrating the proposed algorithm's performance and reliability under various environmental conditions. Comparative analyses with state-of-the-art methods revealed the algorithm's superiority in terms of accuracy and robustness, even in challenging scenarios with sparse AP distributions or complex signal characteristics.
Through this study, a novel and practical user localization approach is established, highlighting its potential for real-world applications. By leveraging only smartphone sensors, this research offers an innovative and scalable solution for location estimation, paving the way for the broader adoption of smartphone-based localization technologies across diverse domains.한국과학기술원 :건설및환경공학과
지반-객체 상호작용 시뮬레이션을 위한 재료점법의 향상
No full text학위논문(박사) - 한국과학기술원 : 건설및환경공학과, 2025.2,[vii, 105 p. :]Soil-object interactions, such as the penetration of construction machines into soils and the movement of wheels over soil surfaces, are common in many engineering applications. Nevertheless, it remains challenging to simulate these interactions for many reasons including large soil deformations and complex frictional contact conditions.
The material point method (MPM), a hybrid Lagrangian-Eulerian method, has become popular for simulating large deformations of history-dependent materials such as soil. For efficient simulation of soil-object interactions, however, the standard MPM faces three significant challenges: (1) volumetric locking in nearly incompressible soils, (2) complex geometry of objects, and (3) sharp gradients in numerical solutions arising from contact at soil-object interfaces.
To address these challenges, this thesis presents three enhancements to the standard MPM. First, we introduce a new assumed deformation gradient method to circumvent volumetric locking in MPM solutions to nearly incompressible soils. This method calculates the assumed deformation gradient through a volume-averaging operation built on the standard particle--grid transfer scheme in the MPM. As a result, the method is simple and general for various explicit MPM formulations.
Second, we present an approach to incorporating objects with complex geometry into the MPM. This approach leverages the level-set method to efficiently represent complex geometric information. It then couples these level-set objects with material points through an algorithm specifically designed to robustly handle frictional contact.
Third, we propose a mapping method for addressing sharp gradients in MPM solutions. The proposed method alleviates sharp gradients by reparameterizing equations into a parametric domain. As the solutions of the reparameterized equations are free of sharp gradients, they can be calculated with a uniform background grid as in the standard MPM.
Finally, for verification and validation, we apply the enhanced MPM to simulation of a variety of soil-object interaction examples. The results of these examples confirm the accuracy and efficiency of the enhanced MPM simulations.한국과학기술원 :건설및환경공학과
위치특이적 비정규 아미노산 도입 및 생직교화학 기술을 이용한 단백질 변형
No full text학위논문(박사) - 한국과학기술원 : 화학과, 2025.2,[vi, 76 p. :]Proteins are linear polymers of up to few thousands amino acids, which are complex biomacromolecules capable of performing numerous roles within living organisms. Maintaining the functional groups on the protein surface intact while introducing desired functional groups at desired locations could be of great help in the development of novel enzymes and functional proteins. This research focused on expanding the functionality of various proteins, such as CRISPR/Cas9, by integrating site-selective introduction techniques for non-canonical amino acids using genetic code expansion technology and bioorthogonal reaction techniques for these amino acids.한국과학기술원 :화학과
Dopaminergic modulation and dosage effects on brain state dynamics and working memory component processes in Parkinson's disease
No full textParkinson's disease (PD) is primarily diagnosed through its characteristic motor deficits, yet it also encompasses progressive cognitive impairments that profoundly affect quality of life. While dopaminergic medications are routinely prescribed to manage motor symptoms in PD, their influence extends to cognitive functions as well. Here we investigate how dopaminergic medication influences aberrant brain circuit dynamics associated with encoding, maintenance and retrieval working memory (WM) task-phases processes. PD participants, both on and off dopaminergic medication, and healthy controls, performed a Sternberg WM task during fMRI scanning. We employ a Bayesian state-space computational model to delineate brain state dynamics related to different task phases. Importantly, a within-subject design allows us to examine individual differences in the effects of dopaminergic medication on brain circuit dynamics and task performance. We find that dopaminergic medication alters connectivity within prefrontal-basal ganglia-thalamic circuits, with changes correlating with enhanced task performance. Dopaminergic medication also restores engagement of task-phase-specific brain states, enhancing task performance. Critically, we identify an "inverted-U-shaped" relationship between medication dosage, brain state dynamics, and task performance. Our study provides valuable insights into the dynamic neural mechanisms underlying individual differences in dopamine treatment response in PD, paving the way for more personalized therapeutic strategies.