187037 research outputs found
Sort by
대규모 의생명정보 데이타베이스 상의 임상시험 시뮬레이션을 위한 정밀의료 프레임워크의 개발
No full text학위논문(박사) - 한국과학기술원 : 의과학대학원, 2025.2,[iv, 64 p. :]Precision medicine employs genetic, environmental, and lifestyle factors to prevent, diagnose, and treat diseases, thereby transforming the field of medicine. Historically, clinical trials have served as a reliable tool for evaluating the efficacy and safety of medical interventions. As precision medicine advances, researchers are developing various tools, including in silico trials, to address its limitations. This research proposes a precision medicine framework to simulate clinical trials on a large-scale biomedical database. A proof-of-concept study utilizing real-world data showed that non-pharmacological intervention (i.e., physical activity) reduced the risk of self-harm by approximately a fourth in depressed patients with a certain genotype. These findings suggest the potential use of data-driven methodologies to simulate comparative effectiveness studies and uncover novel biomarkers in conjunction with clinical interventions. In conclusion, this research provides insights for efficient clinical research and biomedical knowledge discovery.한국과학기술원 :의과학대학원
픽셀형 감마카메라와 기계학습 알고리즘을 활용한 사용후핵연료 안전조치 반출전 검사 장비 개발에 관한 연구
No full text학위논문(박사) - 한국과학기술원 : 원자력및양자공학과, 2025.2,[v, 82 p. :]The international Atomic Energy Agency (IAEA) implements nuclear safeguards to prevent the diversion of special nuclear materials. Many countries, including the Republic of Korea, are facing imminent saturation of spent fuel storage pools, which is expected to significantly increase the IAEA’s budget and personnel burden due to the surge in transfer verification demands in the near future. Therefore, this study focused on developing a new detector and verification methodology that can effectively alleviate the implementation burden of inspection agencies by providing competitive verification performance compared to existing commercial technologies while reducing personnel and cost burdens. The detector design was optimized using Monte Carlo method-based particle transport simulation code(MCNP), and the minimum level of defects required for identification was quantitatively evaluated. Various supervised learning models and anomaly detection models were assessed for their applicability in determining the presence and location of defects. Notably, to ensure reliability for regulatory implementation, explainable AI techniques were employed to evaluate the interpretability of the model’s predictions regarding defect presence and location.한국과학기술원 :원자력및양자공학과
직접에너지 적층을 이용한 원자로 압력용기 SA508 강의 기계적 특성 평가
No full text학위논문(박사) - 한국과학기술원 : 원자력및양자공학과, 2025.2,[xiv, 123 p. :]Small modular reactors (SMRs) have garnered significant attention as next-generation nuclear power plants due to their advanced safety systems, compact size, and lightweight design, offering versatile applications. Among these, pressurized water reactor (PWR)-type SMRs incorporate reactor pressure vessels (RPVs) within the reactor core. To enhance the advantages of SMRs, the design of RPVs has become increasingly complex, necessitating superior mechanical properties. In this thesis, the feasibility of using Laser Powder Directed Energy Deposition (LPDED), an additive manufacturing technique suitable for fabricating large and complex components, was evaluated for the production of RPV components. It was observed that the microstructural characteristics of LPDED-fabricated samples varied significantly depending on thermal conditions, which in turn greatly influenced their mechanical properties. Through optimization of thermal conditions, LPDED-produced samples were found to exhibit superior room- and high-temperature tensile strength, as well as excellent impact toughness, compared to conventionally manufactured samples, even without post-heat treatment. These findings demonstrate the potential of additive manufacturing for the production of reactor pressure vessel components, highlighting its applicability and effectiveness.한국과학기술원 :원자력및양자공학과
자기 스커미온 및 마그논을 이용한 차세대 컴퓨팅 소자 연구
No full text학위논문(박사) - 한국과학기술원 : 물리학과, 2025.2,[iv, 100 p. :]We investigate next-generation computing devices using magnetic skyrmions and magnons as information carriers. Throughout history, computing devices have been developed using various mechanisms, and now, the distinct spin structures within magnetic materials present a promising new approach. Magnetic skyrmions, with their topological protection and unique dynamic properties, are studied for potential applications in energy-efficient memory and logic devices. Through micromagnetic simulations and experiments, we demonstrate skyrmion-based computing devices, including a skyrmion guiding device for skyrmion racetrack memory and a logic adder device enabled by skyrmion annihilation. Magnons, the quanta of spin waves, provide a foundation for non-volatile and low-power computational devices in both classical and quantum domains. By developing magnonic devices such as spin wave phased array antennas and programmable interference systems, we realize real-time, directional control and coherent manipulation of spin waves, demonstrating the potential of magnons in on-chip data transfer and logic operations. For the quantum magnon devices, we investigate a hybrid magnon-photon system, aiming to realize an on-chip, superconducting quantum circuit-compatible hybrid magnonic system. Utilizing YSGG/YIG thin films on a superconducting NbN platform integrated by the flip-chip technique, we achieve on-chip strong magnon-photon coupling with clear band anti-crossing, providing a reliable platform for exploring magnon dynamics for potential quantum applications. The outcomes of this research not only demonstrate effective applications of skyrmion and magnon devices but also establish a foundation for future advancements in hybrid quantum magnonics, paving the way toward scalable, energy-efficient architectures for next-generation computing devices.한국과학기술원 :물리학과
합성 반강자성체에서의 비선형 마그논 현상에 관한 연구
No full text학위논문(박사) - 한국과학기술원 : 물리학과, 2025.2,[iii, 70 p. :]In spintronics, it is important to find low energy controllable and highly stable information carrier. The magnon, often called the quantum of spin-wave (SW) when collective magnon excitation occurs, is especially the promising candidate as highly stable information carrier. Wave properties, such as interference, superposition, dispersive behaviors, of the magnon occupy unique position in the field of spintronics, which invokes new subfield of spintronics ‘Magnonics’. However, there is a critical problem in the magnonics, that is the nonlinearity. The nonlinear phenomenon of magnon is intrinsic and it becomes dominant when the excitation power is high or bias field is low. Although it is difficult to analyze this effect and to imagine intuitive mechanism in the view of physics framework, it is inevitable to confront the effect. Particularly the nonlinear effect on antiferromagnetic spin waves and various hybrid magnonic system is missing. In this thesis, at firstly we investigate nonlinear spin-wave phenomena in the hybrid magnonic system with magnon-magnon coupling. We use synthetic antiferromagnet where the spins are ferromagnetically coupled in intralayer while antiferromagnetically coupled in the interlayer. The eigenmodes of spin-waves in the synthetic antiferromagnet (optic/acoustic modes) interact nonlinearly around the anti-crossing gap between the two modes. As we increase the RF excitation power, it is observed mode hopping which is order of almost 5GHz-jump between acoustic and optic magnons due to three-magnon process with multi-band scattering among propagating and Kittel magnons. Besides, magnonic band of this system reveals hysteretic behavior with respect to field sweep direction. This suggests magnon-magnon hybrid coupling can be tuned by external RF-power. Our work indicates possible application in nonlinear frequency shifter, mode decoupler of magnon logic devices which are prominent for modern neuromorphic computing and artificial intelligence (AI).한국과학기술원 :물리학과
금융 시장의 이상 현상 및 글로벌 시장 수익률 예측 변수에 관한 연구
No full text학위논문(박사) - 한국과학기술원 : 경영공학부, 2025.2,[vi, 196 p. :]This dissertation comprises three essays exploring financial market anomalies and a global market return predictor. The first essay introduces a novel measure of Short-Term Overreaction (STO) using weighted daily signed volume to predict stock returns. The findings demonstrate that STO is a significant predictor of subsequent stock returns, offering predictive power beyond the traditional short-term return reversal. STO negatively predicts abnormal returns around earnings announcements, suggesting investor overreaction in high and low STO stocks. The predictability is stronger when investor sentiment is high and for small, illiquid stocks. Interestingly, the strategy of buying low STO stocks and selling high STO stocks is more profitable when returns are value-weighted than equally weighted, primarily due to institutional investors exacerbating overvaluation in high STO stocks.
The second essay examines the salience theory of stock returns, focusing on reference-dependent preferences. Drawing on mental accounting and prospect theory, the study reveals that investors are influenced by their past gains or losses when selecting stocks with high or low salience. Investors experiencing losses tend to overvalue highly salient stocks, while those with gains undervalue low-salience stocks. Empirical evidence shows a pronounced salience effect in loss regions but a muted or reversed effect in gain regions. These findings are pronounced in stocks with low institutional ownership or high-to-arbitrage. The results are also pronounced under high investor sentiment states or high market volatility, and uncertainty states. The results are further validated in G7 developed countries.
The third essay investigates the predictive role of the relative value of gold, specifically the Gold-Copper ratio, on global stock markets. Gold, serving as both a commodity and a safe-haven asset, provides unique insights into investor expectations and market conditions. The study finds that the Gold-Copper ratio has robust predictive power for short-term stock returns in both developed and global markets, particularly during recessions, aligning with gold's role as a safe haven. The ratio also maintains its predictive strength during economic expansions, aligning with Copper’s role as Dr.Copper. Our research emphasizes the predictive power of Gold-Copper ratio with its utility across various economic scenarios. Additional analyses, including robustness checks, out-of-sample tests, and economic significance tests, confirm the consistent and significant predictive capability of the Gold-Copper ratio, highlighting its potential for substantial economic gains.한국과학기술원 :경영공학부
공중 로봇의 자율주행을 위한 복잡한 환경에서의 통합 모션 계획 프레임워크 연구
No full text학위논문(박사) - 한국과학기술원 : 항공우주공학과, 2025.2,[vii, 97 p. :]This study proposes an efficient solution for autonomous flight for aerial robots in complex environments by introducing MAT-CMPC-based global/local planning and reinforcement learning (RL)-based reactive obstacle avoidance. The framework integrates three key components. First, a Medial Axis Transformation (MAT)-based Safe Flight Corridor (SFC) defines safe zones by generating sequential spherical volumes, ensuring safe navigation through global planning. Second, the MAT-Convex Model Predictive Control (CMPC) formulation optimizes local trajectories, ensuring smooth and collision-free navigation. Third, the RL-based reactive obstacle avoidance method enables real-time handling of unforeseen obstacles through quick and reactive control. Simulations demonstrated that the MAT-CMPC framework efficiently navigates complex environments, outperforming existing methods in computation time, trajectory length, and flight duration. Additionally, the RL-based reactive obstacle avoidance method showed adaptability to various obstacle environments, forward velocities, and local planner configurations. Real-world experiments further confirmed the computational efficiency and practical applicability of the RL-based reactive obstacle avoidance method using a multirotor equipped with an onboard computer.한국과학기술원 :항공우주공학과
대규모 실세계 그래프와 텐서 압축 기법
No full text학위논문(박사) - 한국과학기술원 : 김재철AI대학원, 2025.2,[vi, 99 p. :]We live in the `era of big data', where large-scale datasets are continuously expanding, particularly in the forms of graphs representing relationships among various objects and tensors as multi-dimensional arrays. These datasets, containing billions of nodes or entries and rapidly increasing in size, pose significant costs in data storage and transmission when used in their raw form. To address this challenge, we aim to develop methods that efficiently compress large-scale graphs and tensors to optimize data storage and transmission. Due to the unique structural characteristics of graphs and tensors, conventional matrix algorithms face limitations. By considering these structural properties, this dissertation develops efficient compression methodologies across various graph and tensor-related scenarios. Firstly, we present MoSSo, a lossless compression method for dynamic graphs. This method employs a lossless graph summarization method, which is capable of efficient query processing, and introduces the approach to dynamically update summarization results while maintaining the performance of existing static graph summarization methods. Secondly, we propose NeuKron, a constant-size lossy compression method for sparse matrices and tensors. By introducing a recurrent neural network-based deep learning model and efficient reordering techniques for sparse reorderable matrix and tensor, our method outperforms state-of-the-art competitors in both compression size and accuracy. Thirdly, we present ELiCiT, a lightweight deep-learning-based lossy compression method for tensors without input data assumption, improving upon the slow compression times of existing deep learning-based approaches while offering enhanced compression performance. This method was successfully applied to various applications, including neural network compression and matrix completion. Lastly, we extended this methodology to develop a constant-size compression model for dynamic sparse matrices and tensors, further broadening its applicability.한국과학기술원 :김재철AI대학원
특징 기여도 분석을 통한 심층신경망 성능 향상 : 엣지 모델에서 거대 기반 모델까지
No full text학위논문(박사) - 한국과학기술원 : 김재철AI대학원, 2025.2,[xi, 112 p. :]Feature attribution is an effective technique in interpreting deep neural networks across a wide range of scales, from compressed convolutional neural networks on edge devices to cloud-based large language models. While feature attribution methods are traditionally used for explanation, this dissertation explores the potential of feature attribution for identifying and addressing model vulnerabilities, ultimately improving model performance and reliability. The study progressively expands its model scale, deriving appropriate techniques for each scale. At the smallest scale, this dissertation identifies that while prediction accuracy is preserved during model compression, attribution map integrity is significantly compromised. To address this, we propose an attribution map matching loss function that simultaneously enhances both model explainability and accuracy. For mid-scale applications, we develop a novel tabular data oversampling framework using transformer-based language models. This approach leverages column-wise self-attention attribution to identify class-representative features, enabling targeted column imputation for generating high-quality synthetic samples. The dissertation then expands to large-scale language models, exploring both input and output token-level attribution. From the input perspective, we propose an attribution-guided key-value cache compression methodology that identifies and differentially preserves crucial information, optimizing memory usage while maintaining model performance. From the output perspective, we develop a token-level importance-based verifier that enhances mathematical reasoning capabilities in large language models through fine-grained supervision signals.한국과학기술원 :김재철AI대학원
기하학적 점들의 배열에 관한 확률
No full text학위논문(박사) - 한국과학기술원 : 바이오및뇌공학과, 2025.2,[iii, [100] p. :]This thesis addresses a problem at the foundations of probability theory and geometry that may help achieve the ideal of “inference from data alone,” while also shedding light on the physical basis of observation and evidence. We assume that there exists a configuration of N geometric points (or particles) that are distinguished from one another only by location. The configuration of a subset of these is known. Given the known configuration, what is the unknown configuration? For example, given 2 points (2 ends of a line segment), where is a 3rd point? Or equivalently, what is the triangle formed by these 3 points? This is a problem of trigonometry with incomplete information. We call it the “N-Point Problem.”
This problem is uninteresting if one makes the common assumption that the space of possibilities is “absolute,” as Newton did. An absolute space is fixed and immutable, with no dependence on whatever may be in it. Therefore the location of each point must be independent of others, so that one configuration has no information about another. However, the existence of absolute space is doubtful, as famously pointed out by Leibniz and Mach. Its existence is asserted rather than logically deduced, in contrast to the space of possibilities in classic problems of statistics. For example, we deduce all possible combinations of cards from only the assumption that there exists a standard deck of 52 cards. We do not directly assume anything about space itself.
This thesis demonstrates that a similar approach can be taken for the N points. Assuming only their existence, we can deduce all possible configurations. Logic implies that the N points will form a N-1-dimensional simplex (a triangle is a 2-simplex, and a tetrahedron a 3-simplex). Logic also requires that all points must be treated equally. We show that this implies circular symmetry around the arithmetic mean location (the centroid of the simplex), which is sufficient to uniquely determine the solution. A separate line of reasoning is based on scale invariance, and results in the same solution. Thus each of these symmetries is sufficient to deduce the other.
Our solution to the N-point problem exemplifies and reinforces the objectivist Bayesian ideal in which probability is an objectively correct measure of evidence (rational belief). Its most obvious applications are in physics and chemistry. However, a more novel application concerns the physical basis of observation. In this model an observer is an internal and present configuration of particles that infers an external or future configuration.한국과학기술원 :바이오및뇌공학과