187037 research outputs found
Sort by
REGULARIZATION BY TEXTS FOR LATENT DIFFUSION INVERSE SOLVERS
No full textThe recent development of diffusion models has led to significant progress in solving inverse problems by leveraging these models as powerful generative priors. However, challenges persist due to the ill-posed nature of such problems, often arising from ambiguities in measurements or intrinsic system symmetries. To address this, here we introduce a novel latent diffusion inverse solver, regularization by text (TReg), inspired by the human ability to resolve visual ambiguities through perceptual biases. TReg integrates textual descriptions of preconceptions about the solution during reverse diffusion sampling, dynamically reinforcing these descriptions through null-text optimization, which we refer to as adaptive negation. Our comprehensive experimental results demonstrate that TReg effectively mitigates ambiguity in inverse problems, improving both accuracy and efficiency
LDMol: Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Space Surpass the AR Models
No full text
최적화 기법과 인공지능을 활용한 나노포토닉스 소자의 지능적 설계
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[vi, 94 p. :]Nanophotonics, which explores light-matter interactions at the nanoscale, has driven significant advancements across various research fields. A primary goal in this domain is to design ultra-compact, high-performance nanophotonic devices that can enable the next generation of photonics. While traditional brute-force, intuition-based forward design methods have yielded successful solutions over the past decades, recent advancements in optimization methods and artificial intelligence present new opportunities to extend these capabilities. This study explores recent progress in the inverse design of nanophotonic devices, where AI and optimization methods are used to automate and improve the design process.
This dissertation first reviews representative optimization methods commonly used in inverse design, including various meta-heuristic algorithms such as trajectory-based, evolutionary, and swarm-based approaches. Additionally, we explore state-of-the-art deep learning techniques, involving discriminative models, generative models, and reinforcement learning. We also introduce and categorize several noteworthy inverse-designed nanophotonic devices and their design methodologies. Next, several nanophotonic devices are inverse-designed by leveraging intelligent algorithms. The first study uses a meta-heuristic optimization technique to inverse design the 1×4 photonic power splitters. The second study presents the inverse design of nanophotonic waveguide structures using fully connected neural networks and a time-efficient data preparation process. The third study covers the inverse design of 1D photonic crystal waveguide structures using generative models, incorporating semi-supervised learning to pseudo-label unlabeled data, significantly lowering the computational burden of electromagnetic simulations. Finally, the fourth study explores transfer learning by training a neural network for a 2D photonic crystal waveguide using knowledge from a neural network designed for a 1D photonic crystal waveguide.
Among these studies, the methods vary in their approach to reducing computational complexity and improving the efficiency of the inverse design process, showcasing a range of strategies from meta-heuristic optimization to advanced neural network architectures and transfer learning techniques. These approaches can successfully address the challenges in nanophotonic structure design, paving the way for future technologies.한국과학기술원 :전기및전자공학부
델타 시그마 모듈레이터 코드의 확률밀도함수 변조 기법을 이용한 초저지터 초저스퍼 프랙셔널 주파수 생성기 설계
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[vi, 70 p. :]This paper proposes a low-spur and low-jitter ring oscillator-based fractional-N digital phase-locked loop (DPLL). To achieve ultra-low jitter in a fractional-N PLL, the quantization noise from the delta-sigma modulator is eliminated by a digital-to-time converter (DTC). However, the nonlinearity of the DTC introduces the residual quantization, resulting in fractional spurs, and noise degradation. To address these issues, this paper proposes three key techniques. First, to achieve low fractional spurs, a technique is presented that splits the delta-sigma modulator code into two time-invariant code. This technique exploits the property that time-invariant codes do not generate spurs even after passing through nonlinearity, thus ensuring that no fractional spurs are present in the phase-locked loop output. Second, a technique for modulating the probability density function of the delta-sigma modulator code is introduced to suppress spurs caused by nonlinearities in other circuits within the DPLL. This method modulates the code to satisfy conditions where no spurs are generated, even when the code itself encounters nonlinearity, thereby suppressing fractional spurs throughout the loop. Third, to achieve ultra-low jitter, a technique is proposed for efficiently predistorting the code that controls the DTC. This technique takes advantage of the fact that the nonlinearity of the DTC circuit can be approximated as a polynomial, allowing for power-efficient elimination of nonlinearity compared to conventional structures. The DPLL using all three proposed techniques achieved a fractional spur of –63 dBc and an rms jitter of 365 fs, while consuming only 9.27 mW of power and occupying an area of 0.146 mm².한국과학기술원 :전기및전자공학부
신뢰할 수 있는 테스트 시간 적응을 위한 향상된 자기 학습 기법
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[vi, 82 p. :]The critical challenge of ensuring reliable test-time adaptation (TTA) in increasingly complex real-world scenarios involving significant distribution shifts is addressed in this thesis. TTA adapts pre-trained models to unseen test domains using self-training with unlabeled test data, maintaining performance when test data differ from the training domain. However, existing methods often fail to account for the interplay between input and label distribution shifts or the unique demands of large-scale models like Vision Transformers (ViTs). This thesis proposes a unified framework to resolve these limitations and introduces novel techniques to enhance robustness, efficiency, and adaptability. By advancing the understanding of feature space dynamics, class-wise confusion patterns, and parameter-efficient adaptation strategies, this work represents a significant step forward in TTA research. These contributions not only address the growing complexity of distribution shifts but also provide a foundation for deploying AI models more reliably in dynamic, real-world environments.한국과학기술원 :전기및전자공학부
5G 트랜시버를 위한 저 잡음 광대역 LO 신호 생성 회로설계
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[iv, 68 p. :]In 5G communication systems, unlike previous generations, the objectives include achieving ultra-high data transmission rates exceeding 20 Gbps and extremely low latency of less than 1 ms. Consequently, the importance of designing efficient wideband local oscillator (LO) signal generators has become paramount. LO generators utilized in 5G transceivers must adhere to stringent requirements, such as low phase noise, broad output frequency range, and precise quadrature phase signal generation. This dissertation explores advanced methodologies for the design of LO signal generators in 5G transceivers to meet these exacting standards. First, it introduces a method for correcting quadrature phase errors by evaluating and adjusting the duty cycle in the
quadrature phase signal generator. Second, it presents a low-jitter signal generation technique using a subsampling phase-locked loop (PLL). Finally, the study delves into a cascading architecture for generating output signals over an exceptionally wide frequency range, thereby addressing the diverse and demanding requirements of 5G communication systems.한국과학기술원 :전기및전자공학부
데이터 압축 기반의 메모리 효율적인 대규모 그래프 신경망 가속
No full text학위논문(박사) - 한국과학기술원 : 전기및전자공학부, 2025.2,[v, 62 p. :]Graph Neural Networks (GNNs) are specialized neural network models designed to interpret data structures that focus on relationships between objects in the form of graphs, and they have now become a key area of research within the field of graph representation learning. Due to their advantage of being applicable across a wide range of domains—such as molecular structure analysis and social networks, where data can naturally be represented as graphs, GNNs are increasingly adopted not only by those within the machine learning field but also by researchers across various domains. GNNs are applied to various fields and expanded applications to real-world services. Among these, web services that handle large-scale graph data, such as e-commerce, social media, and financial services, present an ideal application area for GNNs, as these fields contain rich, hidden information that is difficult for humans to analyze. As a result, building solutions that can accelerate the training and inference of GNN models is essential for GNNs to achieve usability and practical implementation.
Since GNNs perform computations by propagating messages according to the irregular and sparse connection patterns of graphs, their operations are memory-intensive rather than compute-intensive. Moreover, when addressing large-scale graphs owned by companies that provide web services, the size of graphs leads to additional memory capacity issues. This results in the need for expensive hardware resources to support GNN training and inference, while at the same time causing inefficiencies in hardware resource utilization in current neural network acceleration systems. Consequently, memory becomes a critical factor in the acceleration and optimization of large-scale GNN computations, and recent studies have been centered on this topic.
This dissertation proposes a memory-efficient GNN processing system that addresses memory challenges through algorithm-hardware co-optimization based on compressed graph data. Previous research on graph data compression has often failed to fully leverage the unique characteristics of GNN models and datasets. Furthermore, such research commonly does not consider implementation at the real-service level. For this reason, this dissertation proposes a compression-based GNN acceleration solution for two distinct scenarios: training and real-time inference.
Firstly, a comprehensive training framework is proposed, co-designing the algorithm and hardware to utilize compressed graphs for memory-efficient GNN training. The proposed compression approach considers both graph and neural network algorithmic aspects for graph data compression while analyzing and exploiting the unique characteristics of the target graphs used in GNNs. Additionally, a GNN computation optimization method aligned with the compressed graph format is introduced, along with a dedicated hardware accelerator architecture to support it. This offers a practical, memory-efficient GNN processing solution for real-world systems. As a result, this approach enables more efficient GNN processing within existing systems without requiring additional hardware. Next, an FPGA-based GNN serving system using vector quantization is proposed to achieve low-latency real-time GNN inference. Through a comprehensive analysis of the end-to-end GNN inference serving system workflow, the preparation step is identified as the primary bottleneck in the inference process. To address this issue, input and hidden embeddings are offloaded to the device memory with compression. The proposed computation method dynamically manages embeddings and eliminates preparation steps. Consequently, this system enables low-latency real-time GNN inference serving.한국과학기술원 :전기및전자공학부
고령화로 인한 공공갈등의 쟁점과 적응적·예측적 거버넌스를 통한 갈등전환에 관한 연구: 한국과 일본의 사례 분석
No full text학위논문(박사) - 한국과학기술원 : 문술미래전략대학원, 2025.2,[ⅳ, 229 p. :]This study structured and comparatively analyzed public conflicts caused by aging populations in South Korea and Japan into five areas: conflicts over the establishment of elderly welfare facilities and funeral facilities, conflicts between younger and older generations regarding employment, conflicts over responses to social insurance depletion such as pension reform, conflicts related to measures addressing concerns of regional depopulation, and conflicts over immigration policies, including the influx of foreigners. Through this analysis, the study identified the complexity and multidimensional characteristics of public conflicts in South Korea and Japan, intertwined with political, economic, social, and cultural factors. This study emphasizes the necessity of adopting the perspective of conflict transformation, which goes beyond the dimensions of conflict coordination and conflict management—where governments seek to mitigate or suppress public conflicts—to fundamentally resolve such conflicts and convert them into opportunities for social change and development. Specifically, the study utilized adaptive governance and anticipatory governance—widely applied governance models in the field of resource governance—as analytical frameworks for conflict transformation in public conflicts. By employing these frameworks, the study analyzed the fundamental causes of and solutions to public conflicts. Adaptive governance provides flexibility in responding to change and uncertainty, enabling conflict transformation through the deep-learning stages of short-term resolution (Single-Loop), structural improvement (Double-Loop), and fundamental change (Triple-Loop). Anticipatory governance supports the prevention of conflicts and the design of long-term, sustainable policies through data-driven scenario planning and real-time feedback systems. Both governance models prioritize the cooperation and participation of citizens and experts, transparent process management and trust, and the development of sustainable alternatives during the policy-making process. When these two governance models are integrated, governments can prevent public conflicts in advance, establish policies before conflicts intensify, and contribute to the formulation of fundamental policies that prevent the recurrence of public conflicts. This study proposes labor reform to address the dual structure and rigidity of the labor market, as well as the use of public pensions to support youth employment and entrepreneurship as key policies for transforming public conflicts arising from aging populations. Such an approach fosters trust among social members, contributes to the prevention and resolution of public conflicts, and ultimately aids in achieving social integration and building a sustainable society.한국과학기술원 :문술미래전략대학원
형상을 넘어서: 동적 그리고 선택적 가열로 기능이 향상된 3D 프린팅 프로토타입
No full text학위논문(박사) - 한국과학기술원 : 산업디자인학과, 2025.2,[vii, 128 p. :]This dissertation presents novel 3D printing systems that enable dynamic and functional objects through selective and dynamic heating techniques. These methods integrate functionalities such as multi-color visual information, dynamic color-changing interfaces, and shape-changing behaviors without requiring electronic assembly. Three systems are introduced: a multi-shade printing method using a single wood-based filament, ThermoPixels for embedding active thermochromic displays on curved surfaces, and ShrinkCells, rigid shape-changing actuators triggered by selective heating. A computational pipeline supports the design and customization of these systems. Additionally, we validated our objectives through expert interviews, which highlighted the importance of iterative prototyping, functionality integration, and enhancing accessibility in 3D printing. These findings and proposed future directions provide valuable insights for advancing functional 3D printing.한국과학기술원 :산업디자인학과