56016 research outputs found
Sort by
Automated Startup Operation for Nuclear Power Plants based on Multi-Agents with Plant Parameter Prediction Model
No full textDepartment of Nuclear Engineeringclos
A Self-supervised Training for Graph Combinatorial Optimization Problems
Get PDFGraduate School of Artificial IntelligenceNeural combinatorial optimization (NCO) has emerged in that it can find near-optimal solution of com- binatorial optimization (CO) problems faster than existing heuristic algorithms thanks to the recent de- velopment of deep learning. Supervised learning is the most natural way to train the deep learning model for CO. However, there are disadvantages in that it is difficult to collect a dataset due to NP-hard char- acteristics of CO problems. To solve the dataset problem, reinforcement learning (RL) has been being used widely in terms of that the objective function can be used as reward signal naturally. But training with RL on CO problem is also hard since it requires huge search space as problem size grows. And the trained models that solve these CO problems must exist as different models suitable for each problem to deal with different markov decision process formulation, size and distribution. Training these models separately according to the situation is time consuming and costly.
Recently in many deep learning research, a pre-trained model is trained first and used for initial start point and fine-tuned to downstream task to address the cost problem. Many self-supervised method has been proposed to train the pre-trained model. It has been shown through many studies that the various self-supervised learning methods are effective in computer vision, natural language processing, and molecule prediction. Graph CO problems are indispensable for real-world industry problems, but not explored yet due to the lack of dataset problem. To train the pre-trained model for the graph CO problems, I propose a novel self-supervised graph neural network training method named pseudo-task for NCO (PTNCO) that approximates greedy algorithm for graph CO problems.
Through this thesis, I first suggest universal node feature design that unifies the node features used for each problem separately so that generating a feasible solution with a single graph neural network. Using the universal node feature design, synthetic random-featured graph can be generated easily and simple but effective pseudo-labeling method for graph CO problems is suggested, based on the randomly generated feature vectors of the synthetic graph. And the proposed PTNCO method is effective for convergence speed and generalization to unseen size, unseen problem and unseen instance when the pre-trained model by PTNCO is used for initial model of successive downstream CO problems via maximum independent set, minimum vertex cover, maximal cut and set cover problem.clos
Cluster-level thyroid malignancy classification via patch-level Papanicolaou staining and refractive index distribution models
Get PDFGraduate School of Artificial IntelligenceRecent advancements in deep-learning methods have revolutionized medical imaging, achieving remarkable classification, segmentation, and detection performance. This thesis explores the utilization of Papanicolaou staining and intrinsic refractive index (RI) as imaging contrasts to classify human thyroid cell clusters and assess their impact on diagnostic accuracy. Using correlative optical diffraction tomography, both color brightfield images from Papanicolaou staining and three-dimensional RI distribution are simultaneously measured in thyroid fine-needle aspiration biopsy (FNAB) specimens. The proposed method aims to classify benign and malignant cell clusters using both Pap stain and RI images and each modality individually. Among 1,535 thyroid cell clusters obtained from 124 patients, the accuracy of classifiers using solely Pap stain images and RI images is 98.0% each, while combining both modalities achieves 100% accuracy. By analyzing the trained model, the Pap stain image primarily utilizes nucleus size as classification information, whereas the RI image incorporates detailed morphological characteristics of the nucleus. This thesis demonstrates that using different imaging modalities as complements can improve thyroid cancer diagnosis by providing robust performances regardless of which machine learning algorithm is employed.clos
CAM-PAR
Get PDFGraduate School of Artificial IntelligenceAs a sub-task of multi-label classification, a pedestrian attribute recognition (PAR) task aims to train a model to detect various attributes for a given image. To achieve better model performance, It is necessary to understand the characteristic of the pedestrian image. Inevitably, most of the pedestrian images have a low resolution because their source is from surveillance cameras, and it is known that some of the pedestrian attributes are highly correlated with each other. To reflect these characteristics, a number of previous methods are proposed. J.Jia et al., propose disentangled attribute feature learning (DAFL) framework for robust training against noisy pedestrian images. DAFL disentangles one-shared encoder feature to attribute specific features using multi-head attention and achieves significant improvements in model performance. But as additional modules are used for disentanglement, the model becomes more complicated. To address this, we propose Class Activation Map guided Pedestrian Attribute Recognition (CAM-PAR) that disentangle features with no need for additional parameters and explore the use of class activation map in multi-label classification domain. On the other hand, other works focus on relations in pedestrian attributes and propose methods that utilize this prior to predicting attributes. But these previous works are limited to modeling pairwise correlation of pedestrian attributes. We propose a Collaborative Filtering for Attribute Recognition (CFAR) module that models correlation of attribute sets using collaborative filtering and utilizes it for attribute prediction. Experiments on PA100K and RAPv1 datasets show that our proposed model surpasses the baseline method and has achieved competitive results against previous state-of-the-art methods.clos
????????? ?????? 3D ?????? ?????? ??? ?????? ?????? ????????????
No full textGraduate School of Artificial Intelligenceclos
Haze Removal for Visibility Improvement With Diverse Environments
No full textDepartment of Electrical EngineeringThe quality of image is easily degraded by the capturing environments: the medium and ambient light. The scene radiance is attenuated by the medium and the scattered ambient light is added to the line-of-sight owing to the light scattering and absorption. We refer the phenomenon as the haze effect and devise the uni???ed image formation model (UIFM) consists of the direct component (DC), the attenuated scene radiance, and the additive component (AC), the scattered light added to the line-of-sight. The haze effect causes the various quality degradation problems such as the low contrast, loss of details, color distortion, blur, low-light, and glow artifacts. As the visibility of image is severely decreased by the quality degradation problems, the computer vision algorithms yield poor performance on hazy image. To improve the visibility, the single image dehazing has been actively studied by estimating the scene radiance from the hazy image. Note that there are diverse environments which cause the haze effect such as the underwater, white hazy, varicolored hazy, and nighttime hazy conditions. In this dissertation, we propose the haze removal with diverse environments to improve the visibility.
We ???rst explicitly de???ne the AC more precisely to develop the dehazing algorithm for underwater conditions. Although the majority of the existing dehazing methods for underwater images consider the backward scattering component only as the AC, we also include the forward scattering component to the AC. Then ???rst, we estimate the transmission map based on the observation that the scene distance is inversely proportional to the geodesic color distance from the background light. Next, we approximate the point spread function in the forward scattering term to estimate the scene radiance. Finally, we obtain the optimal parameters of the UIFM required for transmission estimation and scene radiance restoration by minimizing a cost function composed of the sharpness, information loss, and dark background prior. The experimental results con???rm that the proposed algorithm considerably improves the quality of the estimated transmission maps and restores scene radiance compared with the existing state-of-the-art methods.
There still remain more hazy environments such as the white, varicolored, and nighttime hazy conditions. To tackle them as well, we propose a novel learning-based universal dehazing regardless of the hazy conditions via haze style transfer (HST) without explicit modeling of AC. Inspired by the advanced unpaired image-to-image translation, we transfer the haze style of one hazy image into the other haze style while preserving the scene radiance. Such HST provides the abundant training dataset of various haze styles for universal dehazing. To this end, we ???rst split the haze styles into K domains using the K-means clustering to overcome the ambiguity of the haze style de???nition. Then we sample random haze style vectors to apply the HST on the source hazy image. Next, the haze quality inspection and the reconstruction modules are used for the faithful HST. Finally, the results of HST are used to train the universal dehazing in a semi-supervised learning manner: supervised learning when the scene radiance image exists and unsupervised learning if not. Note that the universal dehazing is actually one special case of HST into no haze style. Motivated by this, we design the haze style modulator (HSM) inspired by UIFM. The HSM conducts HST according to the given haze style vector and outputs the scene radiance only when the zero vector is given. The extensive experiments show that the proposed framework generates the diverse and high-quality hazy images. Moreover, it is applicable for the purpose of the data augmentation. Our universal dehazing successfully improves the visibility of the hazy image regardless of the haze styles, and outperforms the state-of-the art dehazing methods.clos
Oxidative Photocatalysis to Spatiotemporally Control Protein Modification and Cell Death
No full textDepartment of Chemistryclos
Oxygen redox in LiNiO2 cathodes
No full textAssigning oxidation states and understanding the oxygen redox mechanism is crucial for designing superior cathode materials in lithium-ion batteries. The working mechanism of stoichiometric LiNiO2 has been regarded as Ni-dominant redox with partial O contribution through covalent Ni-O bonding for several decades. However, in this issue of Joule, Morris, Grey, and co-workers reported that Ni rarely participates in the redox reaction, and oxygen primarily acts as the redox center through a combination of experimental analysis and computational prediction. Also, the highly reactive singlet O-2 formation mechanism was elucidated. This work provides an opportunity to reassess the current understanding of conventional cathode materials
Emerging industrial clusters of disaster safety industry in Korea
No full textDisaster-related industries have become essential in strengthening both disaster resilience and national competitiveness. For more efficient disaster management, the Korean government widely integrated the disaster industry with the safety industry in 2013, calling it the "disaster safety industry." This study examines the spatial characteristics of the disaster safety industry and its association with regional industries. In emerging industries such as disaster safety industry, there is a scarcity of information regarding intra-industry transactions, and the industry's scope is often vague, thereby restricting comprehensive analysis. To address this issue, we constructed a quasibusiness transaction network that aggregates firm level data to regional level. A correlation analysis using location quotients (LQ) was conducted to determine the relationship with the existing industry. The disaster safety industry network was highly correlated with regional demand. The cluster analysis results showed that four clusters were derived around large cities in the region, which was statistically significant. As a result, these cluster formations were statistically significantly correlated with science- and technology-related industries. Although the disaster safety industry was fostered by the government, we confirmed that technological innovation based on existing industries related to science-based technology can also promote the development of the disaster safety industry
Geometry design of tethered small-molecule acceptor enables highly stable and efficient polymer solar cells
No full textWith the power conversion efficiency of binary polymer solar cells dramatically improved, the thermal stability of the small-molecule acceptors raised the main concerns on the device operating stability. Here, to address this issue, thiophene-dicarboxylate spacer tethered small-molecule acceptors are designed, and their molecular geometries are further regulated via the thiophene-core isomerism engineering, affording dimeric TDY-alpha with a 2, 5-substitution and TDY-beta with 3, 4-substitution on the core. It shows that TDY-alpha processes a higher glass transition temperature, better crystallinity relative to its individual small-molecule acceptor segment and isomeric counterpart of TDY-beta, and amore stablemorphology with the polymer donor. As a result, the TDY-alpha based device delivers a higher device efficiency of 18.1%, and most important, achieves an extrapolated lifetime of about 35000 hours that retaining 80% of their initial efficiency. Our result suggests that with proper geometry design, the tethered small-molecule acceptors can achieve both high device efficiency and operating stability