• KIPS Transactions on Software and Data Engineering
• /
• v.5 no.11
• /
• pp.549-554
• /
• 2016

ConceptNet is a common sense knowledge base which is formed in a semantic graph whose nodes represent concepts and edges show relationships between concepts. As it is difficult to make knowledge base integrity, a knowledge base often suffers from incompleteness problem. Therefore the quality of reasoning performed over such knowledge bases is sometimes unreliable. This work presents neural tensor networks which can alleviate the problem of knowledge bases incompleteness by reasoning new assertions and adding them into ConceptNet. The neural tensor networks are trained with a collection of assertions extracted from ConceptNet. The input of the networks is two concepts, and the output is the confidence score, telling how possible the connection between two concepts is under a specified relationship. The neural tensor networks can expand the usefulness of ConceptNet by increasing the degree of nodes. The accuracy of the neural tensor networks is 87.7% on testing data set. Also the neural tensor networks can predict a new assertion which does not exist in ConceptNet with an accuracy 85.01%.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.8
• /
• pp.319-330
• /
• 2021

Recently, in the field of open domain natural language question answering, multi-task, multi-hop question answering has been studied extensively. In this paper, we propose a novel deep neural network model using hierarchical graphs to answer effectively such multi-task, multi-hop questions. The proposed model extracts different levels of contextual information from multiple paragraphs using hierarchical graphs and graph neural networks, and then utilize them to predict answer type, supporting sentences and answer spans simultaneously. Conducting experiments with the HotpotQA benchmark dataset, we show high performance and positive effects of the proposed model.

• Proceedings of the Korean Society of Precision Engineering Conference
• /
• 1994.10a
• /
• pp.217-222
• /
• 1994

This paper describes the determination of optimal initial billet size for axisymmetric cold forging products using neural networks. The determination of optimal initial billet size is very important in forging design and forming sequence design, because the result of such designs and forming load can be different by variable initial billet sizes. The forming difficulty has been defined as the degree of difficulty in forming by 3 process ' forward extrusion, backward extrusion and upsetting. By neural networks a forming difficulty can be determined with the ratio of shape and forming process. From the graph of maximum, minimum, and average forming difficulties by variable billet sizes, the optimal billet size can be determined. The initial billets of a solid part and a hollow part whichwas determined by this study are compared with the sequence drawing generated by the one of forming sequence design system.

• Annual Conference of KIPS
• /
• 2022.11a
• /
• pp.574-576
• /
• 2022

최근 여러 분야에서 그래프 신경망(graph neural network, GNN)이 활발히 연구되고 있다. 하지만 지금까지 대부분의 GNN 연구는 단일 GNN 모델의 성능을 향상하는 데 집중되었다. 본 논문에서는 앙상블(ensemble) 기법의 대표적 기법인 그래디언트 부스팅(gradient boosting)을 이용하여 GNN의 앙상블 모델을 만드는 방법을 제안한다. 제안 방법은 앞서 만들어진 GNN의 오차를 경사 하강법(gradient descent)을 이용하여 감소시키는 방향으로 다음 GNN을 생성한다. 이 과정을 반복하여 GNN의 최종 앙상블 모델을 얻는다. 실험에서 GNN의 대표적인 모델인 그래프 합성곱 신경망(graph convolutional network, GCN)에 제안 방법을 적용하여 앙상블 모델을 생성한 결과, 단일 GCN 모델에 비해 노드 분류 정확도가 11.3%p까지 증가하였음을 확인하였다.

• Journal of Institute of Control, Robotics and Systems
• /
• v.22 no.6
• /
• pp.477-482
• /
• 2016

This paper presents a simple approximation-based design approach for consensus tracking of heterogeneous second-order nonlinear systems under a directed network. All nonlinearities of followers are assumed to be unknown and non-identical. In the controller design procedure, graph-independent error surfaces are used and an unimplementable intermediate controller for each follower is designed at the first design step. Then, by adding and subtracting a graph-based term at the second step, the actual controller for each follower is designed by using one neural network employed to estimate a lumped and distributed nonlinearity. Therefore, the proposed local controller for each follower has a simpler structure than existing approximation-based consensus tracking controllers for multi-agent systems with unmatched nonlinearities.

• Journal of Information Processing Systems
• /
• v.18 no.2
• /
• pp.188-196
• /
• 2022

Recently, neural architecture search (NAS) has received increasing attention as it can replace human experts in designing the architecture of neural networks for different tasks and has achieved remarkable results in many challenging tasks. In this study, a path-based computation neural architecture encoder (PCE) was proposed. Our PCE first encodes the computation of information on each path in a neural network, and then aggregates the encodings on all paths together through an attention mechanism, simulating the process of information computation along paths in a neural network and encoding the computation on the neural network instead of the structure of the graph, which is more consistent with the computational properties of neural networks. We performed an extensive comparison with eight encoding methods on two commonly used NAS search spaces (NAS-Bench-101 and NAS-Bench-201), which included a comparison of the predictive capabilities of performance predictors and search capabilities based on two search strategies (reinforcement learning-based and Bayesian optimization-based) when equipped with different encoders. Experimental evaluation shows that PCE is an efficient encoding method that effectively ranks and predicts neural architecture performance, thereby improving the search efficiency of neural architectures.

• Electronics and Telecommunications Trends
• /
• v.36 no.3
• /
• pp.106-118
• /
• 2021

The field of brain science (or neuroscience in a broader sense) has inspired researchers in artificial intelligence (AI) for a long time. The outcomes of neuroscience such as Hebb's rule had profound effects on the early AI models, and the models have developed to become the current state-of-the-art artificial neural networks. However, the recent progress in AI led by deep learning architectures is mainly due to elaborate mathematical methods and the rapid growth of computing power rather than neuroscientific inspiration. Meanwhile, major limitations such as opacity, lack of common sense, narrowness, and brittleness have not been thoroughly resolved. To address those problems, many AI researchers turn their attention to neuroscience to get insights and inspirations again. Biologically plausible neural networks, spiking neural networks, and connectome-based networks exemplify such neuroscience-inspired approaches. In addition, the more recent field of brain network analysis is unveiling complex brain mechanisms by handling the brain as dynamic graph models. We argue that the progress toward the human-level AI, which is the goal of AI, can be accelerated by leveraging the novel findings of the human brain network.

• The Journal of the Convergence on Culture Technology
• /
• v.10 no.1
• /
• pp.471-476
• /
• 2024

With the spread of various smart devices and computing devices, big data generation is occurring widely. Machine learning is an algorithm that performs reasoning by learning data patterns. Among the various machine learning algorithms, the algorithm that attracts attention is deep learning based on neural networks. Deep learning is achieving rapid performance improvement with the release of various applications. Recently, among deep learning algorithms, attempts to analyze data using graph structures are increasing. In this study, we present a graph generation method for transferring to a deep learning network. This paper proposes a method of generalizing node properties and edge weights in the graph generation process and converting them into a structure for deep learning input by presenting a matricization We present a method of applying a linear transformation matrix that can preserve attribute and weight information in the graph generation process. Finally, we present a deep learning input structure of a general graph and present an approach for performance analysis.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.5
• /
• pp.205-212
• /
• 2019

Scene graph is a kind of knowledge graph that represents both objects and their relationships found in a image. This paper proposes a 3D scene graph generation model for three-dimensional indoor environments. An 3D scene graph includes not only object types, their positions and attributes, but also three-dimensional spatial relationships between them, An 3D scene graph can be viewed as a prior knowledge base describing the given environment within that the agent will be deployed later. Therefore, 3D scene graphs can be used in many useful applications, such as visual question answering (VQA) and service robots. This proposed 3D scene graph generation model consists of four sub-networks: object detection network (ObjNet), attribute prediction network (AttNet), transfer network (TransNet), relationship prediction network (RelNet). Conducting several experiments with 3D simulated indoor environments provided by AI2-THOR, we confirmed that the proposed model shows high performance.

• Journal of Intelligence and Information Systems
• /
• v.29 no.3
• /
• pp.419-437
• /
• 2023

As the number of thin filers in Korea surpasses 12 million, there is a growing interest in enhancing the accuracy of assessing their credit default risk to generate additional revenue. Specifically, researchers are actively pursuing the development of default prediction models using machine learning and deep learning algorithms, in contrast to traditional statistical default prediction methods, which struggle to capture nonlinearity. Among these efforts, Graph Neural Network (GNN) architecture is noteworthy for predicting default in situations with limited data on thin filers. This is due to their ability to incorporate network information between borrowers alongside conventional credit-related data. However, prior research employing graph neural networks has faced limitations in effectively handling diverse categorical variables present in credit information. In this study, we introduce the Transformer embedded Graph Convolutional Network (TeGCN), which aims to address these limitations and enable effective default prediction for thin filers. TeGCN combines the TabTransformer, capable of extracting contextual information from categorical variables, with the Graph Convolutional Network, which captures network information between borrowers. Our TeGCN model surpasses the baseline model's performance across both the general borrower dataset and the thin filer dataset. Specially, our model performs outstanding results in thin filer default prediction. This study achieves high default prediction accuracy by a model structure tailored to characteristics of credit information containing numerous categorical variables, especially in the context of thin filers with limited data. Our study can contribute to resolving the financial exclusion issues faced by thin filers and facilitate additional revenue within the financial industry.