• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.127-128
• /
• 2020

본 논문은 딥러닝 네트워크의 압축을 위한 양자화 오프셋의 바이어스 기법을 제안한다. 양자화는 32비트 정밀도를 갖는 가중치와 활성화 데이터를 특정 비트 이하의 정수로 압축한다. 양자화는 원 데이터에 스케일과 오프셋을 더함으로써 수행되므로 오프셋을 위한 합성곱 연산이 추가된다. 본 논문에서는 입력 활성화 데이터의 양자화 오프셋과 가중치의 합성곱의 출력은 바이어스에 임베딩될 수 있음을 보여준다. 이를 통해 추론 과정 중 오프셋의 합성곱 연산을 제거할 수 있다. 실험 결과는 오프셋의 합성곱이 바이어스에 임베딩이 되더라도 영상 분류 정확도에 영향이 거의 없음을 증명한다.

• Convergence Security Journal
• /
• v.22 no.1
• /
• pp.55-60
• /
• 2022

This study is to analyze trends and relational structures in the energy sector related to renewable energy. For this reason, in this study, we focused on big data including SNS data. SNS utilizes the Instagram platform to collect renewable energy hash tags and use them as a word embedding method for big data analysis and social network analysis, and based on the results derived from this research, it will be used for the development of the renewable energy industry. It is expected that it can be utilized.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.18 no.9
• /
• pp.2464-2482
• /
• 2024

The semantic understanding of numbers requires association with context. However, powerful neural networks overfit spurious correlations between context and numbers in training corpus can lead to the occurrence of contextual bias, which may affect the network's accurate estimation of number magnitude when making inferences in real-world data. To investigate the resilience of current methodologies against contextual bias, we introduce a novel out-of-distribution (OOD) numerical question-answering (QA) dataset that features specific correlations between context and numbers in the training data, which are not present in the OOD test data. We evaluate the robustness of different numerical encoding and decoding methods when confronted with contextual bias on this dataset. Our findings indicate that encoding methods incorporating more detailed digit information exhibit greater resilience against contextual bias. Inspired by this finding, we propose a digit-aware position embedding strategy, and the experimental results demonstrate that this strategy is highly effective in improving the robustness of neural networks against contextual bias.

• KIPS Transactions on Software and Data Engineering
• /
• v.6 no.4
• /
• pp.203-210
• /
• 2017

In this paper, we propose an effective neural network model for image caption generation and model transfer. This model is a kind of multi-modal recurrent neural network models. It consists of five distinct layers: a convolution neural network layer for extracting visual information from images, an embedding layer for converting each word into a low dimensional feature, a recurrent neural network layer for learning caption sentence structure, and a multi-modal layer for combining visual and language information. In this model, the recurrent neural network layer is constructed by LSTM units, which are well known to be effective for learning and transferring sequence patterns. Moreover, this model has a unique structure in which the output of the convolution neural network layer is linked not only to the input of the initial state of the recurrent neural network layer but also to the input of the multimodal layer, in order to make use of visual information extracted from the image at each recurrent step for generating the corresponding textual caption. Through various comparative experiments using open data sets such as Flickr8k, Flickr30k, and MSCOCO, we demonstrated the proposed multimodal recurrent neural network model has high performance in terms of caption accuracy and model transfer effect.

• Journal of the Korean association of regional geographers
• /
• v.21 no.1
• /
• pp.1-20
• /
• 2015

This paper is to provide some suggestions to complement and extend theory of network city, and to consider preliminarily its applicability for development of the Yeongnam region, exploring its normative implications for urban and regional policy and its significance of empirical research. In order to resolve some limitations and problems of network city theory and of empirical research, we need to reconsider systematically analysis methods, to extend indices of connectivity, to reconfirm normative characters inherent in network city theory, to suggest the constitution of cooperative governance, and to develop policies for embedding functional connectivity into internal community. In a preliminary analysis of Yeongnam region on the basis of network city theory, it is not clear whether the urban system of the region is entirely a type of network city, even though it seems to be close to network city. However, in order for the Yeongnam region to orient towards network city, we can point out importance of policy issues such as expansion of transportation and communication infrastructure, strengthening of economic connectivity, constitution of cooperative governance, and local embeddedness of functional network within the region.

• Journal of the Korea Computer Graphics Society
• /
• v.27 no.3
• /
• pp.65-74
• /
• 2021

3D reconstruction means that reconstructing the 3D shape of the object in an image and a video. We proposed a progressive occupancy network architecture that can recover not only the overall shape of the object but also the local details. Unlike the original occupancy network, which uses a feature vector embedding information of the whole image, we extract and utilize the different levels of image features depending on the receptive field size. We also propose a novel network architecture that applies the image features sequentially to the decoder blocks in the decoder and improves the quality of the reconstructed 3D shape progressively. In addition, we design a novel decoder block structure that combines the different levels of image features properly and uses them for updating the input point feature. We trained our progressive occupancy network with ShapeNet. We compare its representation power with two prior methods, including prior occupancy network(ONet) and the recent work(DISN) that used different levels of image features like ours. From the perspective of evaluation metrics, our network shows better performance than ONet for all the metrics, and it achieved a little better or a compatible score with DISN. For visualization results, we found that our method successfully reconstructs the local details that ONet misses. Also, compare with DISN that fails to reconstruct the thin parts or occluded parts of the object, our progressive occupancy network successfully catches the parts. These results validate the usefulness of the proposed network architecture.

• 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.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.12 no.11
• /
• pp.2105-2110
• /
• 2008

Embedding the compact wearable units to monitor the health status of a person has been analysed as a convenient solution for the home health care. This paper presents a method to detect fall from the other activities of daily living and also to classify those activities. This kind of ambulatory monitoring of the elderly and people with limited mobility can not only provide their general health status but also alarms whenever an emergency such as fall or gait has been occurred and a help is needed. A timely assistance in such a situation can reduce the loss of life. This work shows a detailed analysis of the data received from a chest worn sensor unit embedding a 3-axis accelerometer and depicts which features are important for the classification of human activities. How to arrange and reduce the features to a new feature set so that it can be classified using a simple classifier and also improving the classification resolution. Principal component analysis (PCA) has been used for modifying the feature set and afterwards for reducing the size of the same. Finally a Neural network classifier has been used to analyse the classification accuracies. The accuracy for detection of fall events was found to be 86%. The overall accuracy for the classification of Activities or daily living (ADL) and fall was around 94%.

• Knowledge Management Research
• /
• v.22 no.4
• /
• pp.27-40
• /
• 2021

Click-Through Rate (CTR) prediction is a key function that determines the ranking of candidate items in the recommendation system and recommends high-ranking items to reduce customer information overload and achieve profit maximization through sales promotion. The fields of natural language processing and image classification are achieving remarkable growth through the use of deep neural networks. Recently, a transformer model based on an attention mechanism, differentiated from the mainstream models in the fields of natural language processing and image classification, has been proposed to achieve state-of-the-art in this field. In this study, we present a method for improving the performance of a transformer model for CTR prediction. In order to analyze the effect of discrete and categorical CTR data characteristics different from natural language and image data on performance, experiments on embedding regularization and transformer normalization are performed. According to the experimental results, it was confirmed that the prediction performance of the transformer was significantly improved when the L2 generalization was applied in the embedding process for CTR data input processing and when batch normalization was applied instead of layer normalization, which is the default regularization method, to the transformer model.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.8
• /
• pp.355-364
• /
• 2023

As advanced cyber threats continue to increase in recent years, it is difficult to detect new types of cyber attacks with existing pattern or signature-based intrusion detection method. Therefore, research on anomaly detection methods using data learning-based artificial intelligence technology is increasing. In addition, supervised learning-based anomaly detection methods are difficult to use in real environments because they require sufficient labeled data for learning. Research on an unsupervised learning-based method that learns from normal data and detects an anomaly by finding a pattern in the data itself has been actively conducted. Therefore, this study aims to extract a latent vector that preserves useful sequence information from sequence log data and develop an anomaly detection learning model using the extracted latent vector. Word2Vec was used to create a dense vector representation corresponding to the characteristics of each sequence, and an unsupervised autoencoder was developed to extract latent vectors from sequence data expressed as dense vectors. The developed autoencoder model is a recurrent neural network GRU (Gated Recurrent Unit) based denoising autoencoder suitable for sequence data, a one-dimensional convolutional neural network-based autoencoder to solve the limited short-term memory problem that GRU can have, and an autoencoder combining GRU and one-dimensional convolution was used. The data used in the experiment is time-series-based NGIDS (Next Generation IDS Dataset) data, and as a result of the experiment, an autoencoder that combines GRU and one-dimensional convolution is better than a model using a GRU-based autoencoder or a one-dimensional convolution-based autoencoder. It was efficient in terms of learning time for extracting useful latent patterns from training data, and showed stable performance with smaller fluctuations in anomaly detection performance.