187037 research outputs found
Sort by
양자 커널 기반의 이진 스펙트럴 클러스터링 변분 근사법
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iii, 32 p. :]We addresses a variational quantum machine learning method capable of efficiently approximating unsupervised learning for large-scale data. Quantum machine learning can efficiently handle exponentially growing data by embedding it into quantum systems. However, when using variational quantum circuits with fewer degrees of freedom than the quantum system's dimensionality, performance often degrades. Moreover, learning using variational quantum circuits inevitably introduces local minima, regardless of the convexity of the original problem. To address these issues, we propose a spectral clustering method based on a geometric quantum neural network (GQNN) structure with perturbations and kernel principal component analysis (PCA) equations weighted by a kernel, and examine its performance. The proposed method has been shown through numerical methods to approximate the optimal solution stably.한국과학기술원 :전기및전자공학부
교차 시점 지리 이미지 검색 시 시점 차이 최소화를 위한 대조 학습 프레임워크
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iv, 24 p. :]Cross-view geo-image retrieval (CVIR) is a core task within cross-view geo-localization (CVGL), aiming to retrieve a geo-tagged satellite image given a ground-view query image. CVIR faces substantial challenges due to the significant viewpoint gap between satellite and ground views. While geometric transformations can help reduce this gap, they often introduce distortion and limit cross-view understanding, necessitating an effective comparison method. Moreover, existing contrastive learning-based CVIR approaches are useful for cross-modality tasks with significant domain gaps but focus narrowly on hard-negative sampling and fail to explicitly address the viewpoint gap in CVIR. To address these challenges, we propose a novel contrastive learning framework that bridges the viewpoint gap in CVIR by integrating geometric transformations and modality learning. The proposed method employs modality learning with two components: self-modality, which compares original and transformed images, and cross-modality, which compares transformed images with cross-view originals. This framework learns enriched features from both domains while effectively reducing the viewpoint gap through geometric transformations. By learning consistent representations across varying views of the same location, our method achieves state-of-the-art performance on CVIR benchmarks, including CVUSA, CVACT, and VIGOR.한국과학기술원 :전기및전자공학부
경부하 효율 향상을 위해 PWM 제어하는 LLC 컨버터의 동기 정류기 제어를 통한 새로운 영전압 스위칭 기법
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iii, 34 p. :]In this paper, a novel synchronous rectifier control technique of PWM controlled LLC converter for improving light load efficiency is proposed. In conventional PWM control of LLC converter, primary-side switches turn-on at the valley point of drain-source voltage. However, as the load current increases, significant switching losses occur due to higher drain-source voltage and switching frequency of primary-side switches. To solve this problem, the proposed control technique achieves zero-voltage switching (ZVS) of the primary-side switches by reverse current build-up through turning on the secondary-side synchronous rectifier prior to the primary-side switches. Therefore, the proposed control reduces switching loss, which enables further decreasing of switching frequency. Operating at lower frequency also reduces turn-off switching loss and core losses in magnetic components significantly. The validity of the proposed control technique was verified with 340-420 VDC input and 48V/1000W output prototype converter.한국과학기술원 :전기및전자공학부
대형 언어 모델의 효율적인 온칩 메모리 사용을 위한 온더플라이 Top-K 샘플링 기법
No full text학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iii, 19 p. :]Large language models (LLMs) have revolutionized natural language processing, but their growing need to support richer linguistic representations has driven significant increases in dictionary sizes. This growth presents critical challenges for sampling, particularly in terms of latency and memory usage, as conventional methods require processing the entire dictionary. To address these issues, we propose OF-TOKS, an on-the-fly Top-K sampling mechanism that integrates sorting operations into the matrix multiplication computation pipeline. Moreover, by focusing on only elements during sorting through an adapted bitonic sorting algorithm, OF-TOKS drastically reduces memory usage and hides sorting latency within computation cycles. Compared to conventional methods, OF-TOKS achieves up to 99.95\% reduction in memory usage and 99.92\% reduction in sampling latency overhead. Additionally, hardware synthesis results demonstrate a 48.37\% reduction in area and a 27.97\% reduction in power consumption. These results establish OF-TOKS as a scalable and resource-efficient solution for effective sampling in large language models.한국과학기술원 :전기및전자공학부
그래픽 얼굴에 표현된 정서 인식에 대한 시선추적 연구
No full text학위논문(석사) - 한국과학기술원 : 문화기술대학원, 2025.2,[iii, 66 p. :]In avatar-mediated communication, individuals convey their emotional states through their avatars' facial expressions and infer others' emotional states from the avatars of others. However, since avatars are artificially generated computer graphic images, unlike real human faces, it is essential to investigate how people perceive the emotions displayed on graphic faces and identify which information processing approach people rely on. In the present study, we aimed to identify the difference in perceptual sensitivity and the approaches to processing visual information for facial emotion recognition (FER) in graphic and human faces. In Study 1, the sensitivity for emotion recognition for four emotions (anger, happiness, sadness, surprise) for each graphic and human face was measured by implementing the method of adjustment. The results showed that people are more sensitive to happinessless sensitive to anger and sadnessequally sensitive to surprise in graphics than human faces. With the results of Study 1, the eye-tracking study was conducted to identify which approach is predominantly used for FER in each of the graphics and human faces. The results indicated that the local feature effect was discovered in both graphic and human faces but the attention effect appeared only in graphic faces. Integrating the behavioral responses into scan path analysis, featural processing is predominantly used for graphic faces, and configural processing is used for human faces as aligned with the previous studies.한국과학기술원 :문화기술대학원
가상현실에서 적응적 재생을 위한 작업중단점과 태스크 중심 장면 그래프 기반 기록
No full text학위논문(석사) - 한국과학기술원 : 문화기술대학원, 2025.2,[iii, 21 p. :]We propose a task-centric scene graph-based Virtual Reality (VR) recording method that automatically generates task segment-level breakpoints for adaptive playback of a goal-oriented spatial video. While prior research has explored user state detection and playback timing for adaptive replay, the VR recording system itself, which includes scene, user-object interaction, and task breakpoints for adaptive playback, was underexplored. Our method encodes user-object and object-object relationships with a task-centric scene graph and automatically generates breakpoints that segment the timeline into meaningful units by comparing scene graphs across frames. Through user studies in VR cooking scenarios, results demonstrated that our algorithm could accurately set task breakpoints compared to users' perceptual responses of where adaptive playback should occur. Our proposed VR recording system is expected to be utilized not only to be applicable to various VR learning scenarios but also to facilitate the easy production of spatial videos that are adaptive to the user’s input and state, such as in interactive storytelling.한국과학기술원 :문화기술대학원
Emergent narrative games with Large Language Model-based interaction and storytelling: An exploratory study on system design and user experience
No full text학위논문(석사) - 한국과학기술원 : 산업디자인학과, 2025.2,[vii, 60 p. :]대규모 언어 모델은 전통적인 알고리즘의 한계를 넘어 디지털 게임의 경험을 새롭게 구성하는 기술적 기반이 된다. 사용자의 자유로운 선택과 그에 따른 개인화된 이야기를 지향하는 창발적 내러티브는 대규모 언어 모델을 통해 전통적인 알고리즘의 한계들을 넘어 구현될 수 있을 것으로 전망된다. 그러나 대규모 언어 모델의 발산적인 특성은 모든 경험을 통제하던 기존 게임 디자인 과정과 충돌할 수 있으며, 이로 인해 기존 게임에서는 발견할 수 없었던 다양한 문제들이 발생할 수 있다. 본 연구는 대규모 언어 모델을 적용해 상호작용과 내러티브를 유연하게 구성하는 게임 시스템을 제안하며, 해당 게임에 대한 사용자 경험을 탐색적으로 조사한다. 이를 위해 디자인 프로토타입 게임을 제작한 뒤, 해당 게임을 바탕으로 유저 스터디를 진행하였다. 정성 분석 결과를 기반으로, 대규모 언어 모델이 게임에 일으키는 문제에 대한 논의와 모델을 게임 시스템에 효과적으로 적용하기 위한 디자인 시사점을 제안하였다.한국과학기술원 :산업디자인학과
(A) Study of trade secret infringement between joint holders of trade secret - Previous discussions and new approaches -
No full text학위논문(석사) - 한국과학기술원 : 지식재산대학원프로그램, 2025.2,[vii, 114 p. :]본 연구는 영업비밀의 공동보유자 1인의 영업비밀의 사용, 공개 행위가 다른 보유자에 대한 관계에서 영업비밀 침해행위에 해당하는지 여부를 고찰한다. 이 주제에 관한 종래 법원 판결례 4건과 관련 연구들을 정리한 후 기존 논의에서 나타난 이론적 구성과 설명은 타당하지 않다는 점을 논증한다.
영업비밀 공동보유 관계에 관하여 부정경쟁방지법은 침묵하고 있다. 하지만 법률의 흠결과 법률보충적 법형성의 범위와 한계 관점에서 부정경쟁방지법에 법률의 흠결이 존재하지 않는다는 점을 규명할 것이다. 이에 따라 영업비밀 공동보유자 중 1인의 다른 보유자의 동의 없는 영업비밀의 사용, 공개 행위는 부정경쟁방지법 제2조 제3호 (라)목 영업비밀 침해행위의 해당 여부로 판단하면 충분하고, 결국 그 요건 중에서 ‘부정한 이익을 얻거나 영업비밀 보유자에게 손해를 입힐 목적’의 판단으로 수렴된다.한국과학기술원 :지식재산대학원프로그램
Automatic reference arm path-length adjustment for ultrahigh-speed widefield retinal OCTA
No full text학위논문(석사) - 한국과학기술원 : 기계공학과, 2025.2,[iii, 36 p. :]파장 변환 레이저 기반 초고속 광 간섭 단층 촬영은 비침습적으로 넓은 영역을 빠르게 이미징하여 생체 내나 산업 분야에 널리 활용되며, 특히 안구 망막층 이미징에 많이 쓰인다. 퍼진 펄스 모드 잠금 레이저 기반 시스템은 수 MHz 수준의 높은 반복률로 초고속 망막 OCTA 이미징이 가능하지만, 디지타이저와 광검출기의 전자 신호 검출 성능 한계로 빠른 스캔 시 깊이 범위가 제한되는 문제가 있다. 또한 넓은 범위 스캔 시 안구 곡률로 망막층이 깊이 한계를 벗어나거나 신호 세기가 감소한다. 기존 방식에서는 사전 스캔 과정을 통해 망막층의 곡률에 대응하는 기준 암 광경로 길이 값을 얻은 후 이미징 했으나 준비 시간이 늘어나는 단점이 있었다. 본 연구는 해당 방식을 개선해서 스캔 중 자동으로 광경로 길이를 조정하는 방식을 제안하여 사전 준비를 줄이고, 임상 환경에서 피험자의 촬영·대기 시간을 단축하는 데 기여한다.한국과학기술원 :기계공학과
비접촉식 햅틱 인터페이스의 3차원 촉감 재현을 위한 사용자 인지 특성 기반 초음파 구동기 배치 최적화 연구
No full text학위논문(석사) - 한국과학기술원 : 기계공학과, 2025.2,[v, 54 p. :]In this research, proposing a method for arranging elements that minimizes the creation of grating lobes and provides a wide workspace suitable for the use of both hands, utilizing ultrasound-based non-contact haptic technology. The basic principles of ultrasound-based airborne haptic technology and non-contact haptic devices are introduced, and the skin perception threshold is analyzed through acoustic radiation pressure. Based on this, the modulation frequency affecting the user's perception threshold is experimentally evaluated. Additionally, the changes in sound pressure distribution due to the arrangement of commercial transducer arrays are examined through simulations. A method for optimizing the performance of the workspace area using a genetic algorithm is proposed, and to address the computational complexity of the genetic algorithm with respect to the positions of individual elements, a method is devised to reduce the number of variables by using regularly arranged subarrays, thus suppressing grating lobes. In the future, experimental validation and user evaluation of this technology will be conducted to verify whether the design of the actual hardware can accurately recognize three-dimensional forms. This haptic device is expected to make significant contributions to enhancing interactions in various application areas such as virtual reality (VR) and augmented reality (AR).한국과학기술원 :기계공학과